WO2023029117A1 - Method and apparatus for analyzing alum floc feature by using image recognition technology - Google Patents

Method and apparatus for analyzing alum floc feature by using image recognition technology Download PDF

Info

Publication number
WO2023029117A1
WO2023029117A1 PCT/CN2021/119244 CN2021119244W WO2023029117A1 WO 2023029117 A1 WO2023029117 A1 WO 2023029117A1 CN 2021119244 W CN2021119244 W CN 2021119244W WO 2023029117 A1 WO2023029117 A1 WO 2023029117A1
Authority
WO
WIPO (PCT)
Prior art keywords
alum
flower
flowers
image
model
Prior art date
Application number
PCT/CN2021/119244
Other languages
French (fr)
Chinese (zh)
Inventor
***
夏萍
冼峰
吕玉龙
戴毓文
陆劲蓉
徐鸿凯
徐彦琨
魏宁
张双翼
Original Assignee
上海城市水资源开发利用国家工程中心有限公司
上海城投水务（集团）有限公司制水分公司
上海城投水务(集团)有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 上海城市水资源开发利用国家工程中心有限公司, 上海城投水务（集团）有限公司制水分公司, 上海城投水务(集团)有限公司 filed Critical 上海城市水资源开发利用国家工程中心有限公司
Publication of WO2023029117A1 publication Critical patent/WO2023029117A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Definitions

  • the invention relates to the technical field of image processing, in particular to a method and device for analyzing the characteristics of alum flowers using image recognition technology.
  • An effective coagulation dosing control method can control the coagulant dosing in the sense of the actual optimal dosage, so as to achieve the most ideal effluent water quality with the least consumption of chemicals. Effect.
  • Coagulation dosing is an important part of water treatment, which has the characteristics of large lag and nonlinearity. How to realize the automatic control of coagulation dosing has always been a concern of the water industry. Previous dosing control methods, such as mathematical model method, flowing current method, simulated filter method, etc., have not been promoted and applied on a large scale due to limitations and unreliability of one kind or another.
  • each water plant still checks the effect of dosing according to the flow of water entering the plant, combined with manual observation of the effect of alum flowers and the outlet turbidimeter of the sedimentation tank. It mainly relies on the flow ratio to control the dosage of coagulant, and it is impossible to accurately control the water quality. and detection.
  • a fixed programmed automatic control system with a flow ratio is used. The test results are only representative of the water quality at the moment of sampling.
  • the determined coagulant dosage has problems of discontinuity and hysteresis. It is difficult to realize the coagulant dosage during the operation of the water plant. The optimal control of dosage cannot realize the precise control of coagulant dosage.
  • threshold segmentation and morphological processing can be performed on the collected alum flower images, and the main image features can be selected from many image features for standardization processing.
  • the standardized features Conduct training to judge whether the amount of alum is appropriate.
  • Computer vision realized through deep learning has been able to help people monitor the production environment, identify product defects, and identify potential failures. Therefore, this technology basically has the conditions for alum flower identification in the water production process of water plants, helping to achieve a more accurate image of alum flowers. Sophisticated and intelligent analysis.
  • the traditional alum flower detection algorithm based on machine learning and image processing mainly includes digital image processing module, feature engineering module and machine learning module, as shown in Figure 1.
  • algorithm complexity due to its serial algorithm logic and the need to repeat feature engineering and machine learning until it meets the corresponding scene, the algorithm runs slowly and has a large amount of repetitive work; in terms of robustness, when the scene changes, the algorithm needs to be re-started. Design, algorithm robustness is poor.
  • the purpose of the present invention is to provide a method and device for analyzing the characteristics of alum flowers using image recognition technology, so as to realize the intelligent control of coagulant dosing, so as to realize the precise control of coagulant. add.
  • the present invention proposes a kind of method utilizing image recognition technology to analyze alum flower feature, comprises the steps:
  • Step S1 use the underwater camera to collect alum flower images, mark the collected images, generate training data, establish an AI model for alum flower recognition, use the marking data as a supervisory signal, conduct independent training on the alum flower recognition AI model, and obtain a well-trained Alum flower recognition AI model;
  • Step S2 using the underwater camera to collect images of alum flowers, and using the trained AI model for identifying alum flowers to identify them;
  • step S3 the single frame image is divided into M ⁇ M regions, and the quantitative index of alum flower is calculated for each single area, and the quantitative evaluation index of the single frame image is obtained by combining various quantitative indexes of alum flower in the single area.
  • step S1 further includes:
  • Step S100 using the underwater camera to collect images of alum flowers within its depth of field
  • Step S101 determining the labeling standards for each category of alum flowers, marking the collected images, and generating training data
  • step S102 an AI model for alum flower recognition of a multi-view alum flower detection and segmentation network is established, and the marking data is used as a supervisory signal to train the multi-view alum flower detection and segmentation network to fit the labeled data.
  • step S101 according to the related business of alum flowers and the captured images, determine the labeling standards of the three types of alum flowers, and mark the collected images according to the determined labeling standards. Shaped alum flowers, alum flowers, and fuzzy fluffy images, add note data.
  • the alum flower recognition AI model uses a multi-layer neural network to extract high-level features of the image through multiple layers of convolution and activation layers, and use the extracted high-level features of the image to obtain Get the detection frame of the alum flower, and use the advanced features of the image to classify the type of alum flower through the classification task.
  • step S102 a large amount of alum flower data is used to optimize model parameters through back propagation, and to train the alum flower recognition AI model.
  • the trained alum flower recognition AI model is used to identify and collect images, and attributes such as category, probability and location of each alum flower are detected.
  • the quantitative index of each alum flower category in the single region includes the number of the category in the region, confidence average, confidence median, area average and area median.
  • step S3 the quantitative indicators of three types of alum flowers in all areas of M ⁇ M are combined to obtain 15M 2 -dimensional quantitative indicators of alum flowers, which are the quantitative evaluation indicators of a single frame image.
  • step S3 F frames are selected at equal intervals from the images collected within 1 minute, and the quantitative index of alum flower of a single frame image is calculated, and then the same quantitative index of alum flower in F frames is averaged to obtain the time-quantified alum flower Quantitative indicators.
  • the present invention also provides a device utilizing image recognition technology to analyze the characteristics of alum flowers, comprising:
  • Alum flower recognition AI model construction and training unit used to use underwater camera to collect alum flower images, mark the collected images, generate training data, establish alum flower recognition AI model, use the marking data as a supervisory signal, and identify alum flower
  • the AI model is trained independently to obtain the trained AI model for alum flower recognition
  • the image collection and recognition unit is used to collect the alum flower image by using the underwater camera, and use the trained AI model to identify the alum flower;
  • the quantitative evaluation index detection unit is used to divide a single frame image into M ⁇ M areas, calculate the quantitative index of alum flower for a single area respectively, and combine the three types of alum flower quantitative indicators in a single area to obtain the quantification of a single frame image evaluation index.
  • the quantitative evaluation index detection unit selects F frames at equal intervals from the images collected within 1 minute, calculates the quantitative index of alum flower in a single frame image, and then averages the same quantitative index of alum flower in F frames to obtain the time-quantified Quantitative indicators of alum flowers.
  • the present invention is a method and device for analyzing the characteristics of alum flowers using image recognition technology.
  • an AI model for alum flower recognition the image of alum flowers can be recognized, and the quantitative indicators of alum flowers can be detected, so that it can be judged Whether the amount of alum is appropriate to realize the intelligent control of coagulant dosing, so as to realize the precise dosing of coagulant.
  • Fig. 1 is the schematic diagram of the traditional alum flower detection algorithm based on machine learning and image processing
  • Fig. 2 a is the structural diagram of perceptron (neuron) among the present invention.
  • Figure 2b is a structural diagram of a multi-layer perceptron in the present invention.
  • Fig. 2c is a schematic diagram of the learning (training) process of the multi-layer perceptron in the present invention.
  • Fig. 3 is a flow chart of the steps of a method utilizing image recognition technology to analyze the characteristics of alum flowers of the present invention
  • Fig. 4 is the schematic diagram of raw water purification process
  • Fig. 5 is a schematic diagram of flaky alum flowers, fluffy alum flowers and fuzzy fluffy in the embodiment of the present invention
  • Fig. 6 is a schematic diagram illustrating the depth of field of a simple optical system in an embodiment of the present invention.
  • Fig. 7 is a standard schematic diagram of three types of alum flowers in the embodiment of the present invention.
  • Fig. 8 is the structural representation of the multilayer neural network that alum flower recognition AI model adopts in the specific embodiment of the present invention.
  • FIG. 9 is a schematic diagram of a multi-view alum flower detection and segmentation network in an embodiment of the present invention.
  • FIG. 10 is a schematic diagram of dividing an image in a 2 ⁇ 2 manner for a single frame image in an embodiment of the present invention.
  • Fig. 11 is a schematic diagram of calculation of quantitative index of alum flower in a single region in the embodiment of the present invention.
  • FIG. 12 is a system architecture diagram of a device for analyzing the characteristics of alum flowers using image recognition technology according to the present invention
  • Fig. 13 is a schematic diagram of the importance of various indicators of alum flower quantification for sinking water turbidity in the embodiment of the present invention.
  • Figure 14 is a schematic diagram of the importance of different types of alum flowers in all regions in the embodiment of the present invention.
  • Perceptron (neuron), its structure is as shown in Figure 2a, specifically as follows:
  • the perceptron contains a set of parameters, including linear calculations and nonlinear calculations;
  • a perceptron is a neural network with a single artificial neuron, an input layer, and a set of connections connecting the input unit to the output unit;
  • the goal of the perceptron is to classify the pattern presented to the input unit.
  • the basic operation performed by the output unit is to multiply each input (xn) with its connection strength or weight (wn) and pass the sum of the products to the output unit;
  • the input can be pixels in an image, or more commonly, features extracted from the original image, such as the contours of objects in the image.
  • the perceptron determines whether the image is a member of a class, such as cats.
  • the output can only be one of two states, 'on' if the image is in a category, 'off' otherwise. "On" and "Off" correspond to 1 and 0 in binary values, respectively.
  • Multi-layer perceptron its structure is shown in Figure 2b, specifically as follows:
  • Feedforward network containing multiple layers of perceptrons, layer-by-layer calculation of feedforward network:
  • the input value propagates forward layer by layer from the input layer neurons through the weighted connection, passes through the hidden layer, and finally reaches the output layer to get the output.
  • the weight of the network is fixed, and the state of neurons in each layer only affects the state of neurons in the next layer. The process is as follows:
  • Backpropagation is the process of training perceptron parameters, and the input of backpropagation is the feedforward of forward propagation;
  • the output result is compared with the value given by the trainer, and the difference is used to update the weight of the connected output unit to reduce the error;
  • the weights between the input unit and the hidden layer are updated by backpropagating the error according to how much each weight contributes to the error.
  • the hidden units generate selective features that distinguish different input patterns, so that they can distinguish between different categories in the output layer.
  • Fig. 3 is a flow chart of steps of a method for analyzing the characteristics of alum flowers using image recognition technology in the present invention.
  • a kind of method of utilizing image recognition technology to analyze alum flower feature of the present invention comprises the following steps:
  • Step S1 use the underwater camera to collect alum flower images, mark the collected images, generate training data, establish an AI model for alum flower recognition, and use the marking data as a supervisory signal to independently train the alum flower recognition AI model to obtain training data.
  • the purification of raw water is mainly carried out in flocculation tanks, sedimentation tanks and filter tanks.
  • coagulants are added to the flocculation tanks to form granular crystal alum flowers through chemical reactions, and the water flows through the flocculation tanks into the sedimentation tanks.
  • the alum flowers are basically formed at the water inlet of the sedimentation tank.
  • the granular alum flowers are constantly combined, moving and adsorbing impurities in the raw water in the coagulation tank and the sedimentation tank. The flowers move and absorb each other to form fluffy alum flowers.
  • alum flowers are divided into flaky alum flowers and fluffy alum flowers.
  • flaky alum flowers due to the fixed parameters such as the focal length of the camera, the alum flowers close to the camera will appear blurred on the image, causing the human eye to be unable to distinguish their specific shapes.
  • fuzzy and fluffy for this part of the alum flowers, they cannot be simply classified as flaky and fluffy alum flowers, but It needs to be classified separately as fuzzy and fluffy, as a supplementary category of alum flower. Therefore, considering the business of alum flowers and the influence of cameras, in the AI model of alum flower recognition, alum flowers are divided into: flaky alum flowers, fluffy alum flowers, and fuzzy and fluffy, as shown in Figure 5.
  • step S1 further includes:
  • Step S100 using the underwater camera to collect images of alum flowers within its depth of field.
  • the optical system with positive focal length can realize infinite distance imaging, but it cannot be done due to the energy loss of the optical path and the size of the CCD (optical sensor) phase element. Imaging at infinity to the camera. It is stipulated that the minimum diameter of the diffuse spot supported by the camera is ⁇ , and the depth of field of the optical system of the camera can be calculated through the focal length of the camera, the size of the phase element, and the relevant parameters of the optical system.
  • the physical meaning of the depth of field the object can get a relatively clear image on the image plane (CCD) within the depth of field of the object plane.
  • the shooting range of the underwater camera is the alum flowers in the cuboid near the depth of field.
  • the underwater camera adopts two cameras, which are placed at the water inlet of the sedimentation tank and installed on the three-tube supporting plate, with a distance from the water surface 80cm, shot at an angle of 45 degrees obliquely downward.
  • Step S101 determining the labeling standards for each category of alum flowers, marking the collected images, and generating training data.
  • a certain amount of marking data is required as a supervisory signal, and the alum flower recognition AI model is allowed to train independently and fit the labeled data. Therefore, labeling data is an important part of the alum flower recognition AI model. Combined with the alum flower business and the shooting situation, the alum flower is more fine-grained. In terms of labeling, a problem is introduced: how to distinguish various types of alum flowers according to the captured images.
  • step S102 an AI model for alum flower recognition of a multi-view alum flower detection and segmentation network is established, and the marking data is used as a supervisory signal to train the multi-view alum flower detection and segmentation network to fit the labeled data.
  • the alum flower recognition AI model adopts a multilayer neural network, as shown in Figure 8, in this multilayer neural network, the convolution layer+activation function can be regarded as a perceptron, and the whole network includes Multi-layer perceptron, features are continuously extracted and compressed through multiple convolutional layers + activation functions (perceptrons), and finally relatively high-level features can be obtained, that is, the original features are concentrated step by step through the multi-layer neural network , the final features are more reliable, and finally use the last layer of features to do various tasks: such as classification (such as judging the category), regression (such as returning the detection frame), etc.
  • classification such as judging the category
  • regression such as returning the detection frame
  • the operating principle of the alum flower recognition AI model is mainly:
  • the detection frame (position and size) of the alum flower is obtained by regression
  • the type of alum flower (fluffy, flaky, fuzzy alum flower) can be obtained by classification.
  • the alum flower detection and segmentation algorithm based on deep learning can automatically extract and combine features without manual adjustment of the algorithm.
  • the present invention precipitates detection and segmentation algorithms and industrial vision platforms in actual combat, and introduces practical experience and academic knowledge into the algorithm; in terms of video information, the present invention can improve video information extraction and integration through optical flow method and video spatiotemporal feature extraction technology, Improve the accuracy and robustness of detection and segmentation algorithms.
  • the alum flower detection and segmentation algorithm based on multi-view and deep learning can detect alum flower better under different underwater shooting conditions.
  • Step S2 using the underwater camera to collect images of alum flowers, and using the trained AI model for identifying alum flowers to identify them.
  • the trained alum flower recognition AI model can detect attributes such as category, probability and position of each detected alum flower in the captured image.
  • attributes such as category, probability and position of each detected alum flower in the captured image.
  • Table 1 The related attributes of alum flowers detected by the detection and segmentation network
  • Step S3 divide the single-frame image into M ⁇ M regions, calculate the quantitative index of the alum flower for the single area, and combine the three types of alum flower quantitative indicators in the single area to obtain the quantitative evaluation index of the single-frame image, in which the single
  • the quantitative indicators of each alum flower category in the region include the number of the category in the region, the confidence average, the confidence median, the area average, and the area median.
  • the single frame image was divided into M ⁇ M regions. Taking 2 ⁇ 2 as an example, as shown in Figure 10, the entire image is divided into 4 regions as shown in the figure in the form of 2 ⁇ 2, and then the quantitative index of alum flower is calculated for a single region.
  • Calculation of quantitative indicators of alum flower in single frame and single area take the upper left area of Figure 10 as an example to illustrate the calculation method of quantitative index of alum flower in single frame and single area. As shown in Figure 11, for each type of alum flower category in a single area, the number, confidence average, confidence median, area average, and area median of the category in the area are calculated respectively.
  • c represents the category of alum flowers
  • i represents the division area
  • each index of alum flower represents:
  • Confidence is the probability that the alum flower recognition AI model predicts the detection of alum flower, which represents the confidence that the detected alum flower is the current category;
  • Average area and median area The ratio of the sum of the area of a certain type of alum flowers detected in the current area to the area of the area represents the density of this type of alum flowers in the area.
  • Combination of single-frame and multi-region alum quantitative indicators Combining the three types of alum quantitative indicators in all regions of M ⁇ M can obtain 15M 2- dimensional alum quantitative indicators, which are quantitative evaluation indicators for single-frame images.
  • Time quantification of alum quantitative indicators Since the raw water data is updated once a minute, and due to the FPS characteristics of the camera, there are about 360 frames of images in one minute. If all images participate in the calculation of alum quantitative indicators, it will result in: (1) every frame The images have to go through the calculation of the AI model of alum flower recognition and the quantitative evaluation of multi-region alum flower in a single frame, and the calculation complexity is too large; (2) the quantitative indicators of alum flower in adjacent frames are similar, and the features are redundant.
  • the method adopted is to select F frames at equal intervals from the images collected within 1 minute, and calculate the quantitative index of alum flower for a single frame image. Finally, the F frames are the same The quantitative index of the alum flower is averaged to obtain the quantitative index of the alum flower after time quantification.
  • Fig. 12 is a system architecture diagram of a device for analyzing characteristics of alum flowers using image recognition technology according to the present invention. As shown in Figure 12, a kind of device of the present invention utilizes image recognition technology to analyze the characteristics of alum flowers, comprising:
  • the alum flower recognition AI model construction and training unit 10 is used to collect alum flower images with an underwater camera, mark the collected images, generate training data, establish an alum flower recognition AI model, use the marking data as a supervisory signal, and identify the alum flower Recognize the AI model for autonomous training, and obtain the trained AI model for alum flower recognition.
  • the purification of raw water is mainly carried out in flocculation tanks, sedimentation tanks and filter tanks.
  • Coagulants are added to the flocculation tanks to form granular crystal alum flowers through chemical reactions.
  • the water flows into the sedimentation tanks through the flocculation tanks.
  • the formation of alum flowers is basically completed.
  • the granular alum flowers are constantly combined with each other, moving and adsorbing impurities in the raw water.
  • the shape changes from granular to flake. Fluffy alum flowers, at this time the density of alum flowers is greater than that of water, and finally settle to the bottom of the sedimentation tank due to gravity, completing preliminary water purification and reducing the turbidity of raw water.
  • alum flowers Through the business analysis of alum flowers, according to their shape, they can be divided into flaky alum flowers and fluffy alum flowers. However, due to the fixed parameters such as the focal length of the camera, the alum flowers close to the camera will appear blurred on the image, causing the human eye to be unable to distinguish their specific shapes. For this part of the alum flowers, they cannot be simply classified as flaky and fluffy alum flowers, but It needs to be classified separately as fuzzy and fluffy, as a supplementary category of alum flower. Therefore, considering the business of alum flowers and the influence of cameras, in the AI model of alum flower recognition, alum flowers are divided into: flaky alum flowers, fluffy alum flowers, and fuzzy and fluffy, as shown in Figure 5.
  • alum flower recognition AI model construction and training unit 10 further include:
  • the image acquisition module is used to utilize the underwater camera to collect images of alum flowers within the depth of field.
  • the shooting range of the underwater camera is the alum flowers in the cuboid near the depth of field.
  • the marking module is used to determine the labeling standards of various types of alum flowers, mark the collected images, and generate training data.
  • a certain amount of marking data is required as a supervisory signal, and the alum flower recognition AI model is allowed to train independently and fit the labeled data. Therefore, labeling data is an important part of the alum flower recognition AI model. Combined with the alum flower business and the shooting situation, the alum flower is more fine-grained. In terms of labeling, a problem is introduced: how to distinguish various types of alum flowers according to the captured images
  • the training module is used to establish the alum flower recognition AI model of the multi-view alum flower detection and segmentation network, use the marking data as the supervision signal, train the multi-view alum flower detection and segmentation network, and fit the labeled data.
  • the training module uses a multi-layer neural network to establish an AI model for alum flower recognition.
  • the multi-layer neural network is shown in Figure 8.
  • the convolution layer + activation function can be regarded as a perceptron, and the entire network consists of multiple layers Perceptron, features are continuously extracted and compressed through multiple convolutional layers + activation functions (perceptrons), and finally relatively high-level features can be obtained, that is, the original features are concentrated step by step through the multi-layer neural network, and finally The obtained features are more reliable, and finally the last layer of features can be used to do various tasks: such as classification (such as judging the category), regression (such as returning to the detection frame), etc.
  • the operating principle of the alum flower recognition AI model is mainly:
  • the detection frame (position and size) of the alum flower is obtained by regression
  • the type of alum flower (fluffy, flaky, fuzzy alum flower) can be obtained by classification.
  • the image collection and recognition unit 11 is used to collect images of alum flowers with an underwater camera, and use the trained AI model for recognition of alum flowers to perform recognition.
  • the trained alum flower recognition AI model can detect attributes such as category, probability and position of each detected alum flower in the captured image.
  • attributes such as category, probability and position of each detected alum flower in the captured image.
  • Table 1 The related attributes of alum flowers detected by the detection and segmentation network
  • the quantitative evaluation index detection unit 12 is used to divide the single-frame image into M ⁇ M regions, calculate the alum flower quantification index for the single area respectively, and combine the three types of alum flower quantification indexes in the single area to obtain the single-frame image Quantitative evaluation indicators, wherein the quantitative indicators of each alum flower category in a single area include the number of the category in the area, confidence average, confidence median, area average, and area median.
  • the single frame image was divided into M ⁇ M regions. Taking 2 ⁇ 2 as an example, as shown in Figure 10, the entire image is divided into the following 4 regions in the form of 2 ⁇ 2, and the quantitative indicators of alum flowers are calculated for a single region
  • Calculation of quantitative indicators of alum flower in single frame and single area take the upper left area of Figure 10 as an example to illustrate the calculation method of quantitative index of alum flower in single frame and single area. As shown in Figure 11, for each type of alum flower category in a single area, the number, confidence average, confidence median, area average, and area median of the category in the area are calculated respectively.
  • c represents the category of alum flowers
  • i represents the division area
  • each index of alum flower represents:
  • Confidence is the probability that the alum flower recognition AI model predicts the detection of alum flower, which represents the confidence that the detected alum flower is the current category;
  • Average area and median area The ratio of the sum of the area of a certain type of alum flowers detected in the current area to the area of the area represents the density of this type of alum flowers in the area.
  • Combination of single-frame and multi-region alum quantitative indicators Combining the three types of alum quantitative indicators in all regions of M ⁇ M can obtain 15M 2- dimensional alum quantitative indicators, which are quantitative evaluation indicators for single-frame images.
  • Time quantification of alum quantitative indicators Since the raw water data is updated once a minute, and due to the FPS characteristics of the camera, there are about 360 frames of images in one minute. If all images participate in the calculation of alum quantitative indicators, it will result in: (1) every frame The images have to go through the calculation of the AI model of alum flower recognition and the quantitative evaluation of multi-region alum flower in a single frame, and the calculation complexity is too large; (2) the quantitative indicators of alum flower in adjacent frames are similar, and the features are redundant.
  • the method adopted is to select F frames at equal intervals from the images collected within 1 minute, and calculate the quantitative index of alum flower for a single frame image. Finally, the F frames are the same The quantitative index of the alum flower is averaged to obtain the quantitative index of the alum flower after time quantification.
  • Experimental data Communicating with the business experts of the water plant, because the water purification system of the water plant is undergoing business transformation, the excessive use of coagulant has caused alum flowers at the water inlet of the sedimentation tank, most of which are granular. Therefore, relatively conventional data from April 24th to April 25th were used in the initial experiments to confirm relevant parameters. In this embodiment, 38 hours of data are used as a training set, and 10 hours of data are used as a test set. During the experiment, the training set: 2280 records, and the test set: 600 records.
  • Quantitative index time quantification parameter F frame experiment of alum flower During the experiment, 10, 25, 30, 40, 50, 60, and 70 frames of images were sampled at equal intervals every minute to calculate the quantitative index of alum flower, and a support vector regression model was established ( Only use the quantitative indicators of alum flowers to model), in the same test set, the scores of each model are as follows in Table 3:
  • the establishment of a multimodal data source lag regression model of turbidity of submerged water can also evaluate whether the quantitative indicators of alum flowers are reasonable based on the trained model.
  • the normalized importance of various indicators of alum flower quantification to the turbidity of the sinking water can be calculated, and the quantitative indicators are given, as shown in FIG. 13 .
  • the quantitative indicators (number of alum flowers, confidence average, confidence median, area average, and area median) of each category of alum flowers per minute in all areas are averaged, and displayed on the web interface after normalization. On the histogram, it can be seen from the quantitative importance indicators of the quantitative indicators of alum flowers in each region that the coagulation effect of alum flowers is better when the number of fluffy alum flowers is larger and the area is larger.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Image Analysis (AREA)

Abstract

A method and apparatus for analyzing an alum floc feature by using an image recognition technology. The method comprises the following steps: step S1, acquiring an alum floc image by using an underwater camera, marking the acquired image, generating training data, establishing an alum floc recognition AI model, and performing, with the marking data as a supervision signal, autonomous training on the alum floc recognition AI model to obtain a trained alum floc recognition AI model; step S2, acquiring an alum floc image by using the underwater camera, and performing recognition by using the trained alum floc recognition AI model; and step S3, perform regional division on a single-frame image to form M×M regions, respectively calculating alum floc quantization indexes for single regions, and combining the alum floc quantization indexes of the single regions to obtain a quantization evaluation index of the single-frame image.

Description

一种利用图像识别技术分析矾花特征的方法及装置A method and device for analyzing characteristics of alum flowers using image recognition technology 技术领域technical field
本发明涉及图像处理技术领域，特别是涉及一种利用图像识别技术分析矾花特征的方法及装置。The invention relates to the technical field of image processing, in particular to a method and device for analyzing the characteristics of alum flowers using image recognition technology.
背景技术Background technique
混凝的效果直接影响后续工艺的控制，有效的混凝投药控制方法，可以在实际最佳投药量的意义上控制混凝剂投加，从而达到以最少的药剂消耗获得最理想的出水水质的效果。混凝投药是水处理的一个重要环节，具有大滞后和非线性等特点。如何实现混凝投药的自动控制一直是制水行业关注的一个问题。以往的投药控制方法，如数学模型法、流动电流法、模拟滤池法等都因为存在这样或那样的局限性和不可靠性，而没有形成规模性的推广和应用。The effect of coagulation directly affects the control of subsequent processes. An effective coagulation dosing control method can control the coagulant dosing in the sense of the actual optimal dosage, so as to achieve the most ideal effluent water quality with the least consumption of chemicals. Effect. Coagulation dosing is an important part of water treatment, which has the characteristics of large lag and nonlinearity. How to realize the automatic control of coagulation dosing has always been a concern of the water industry. Previous dosing control methods, such as mathematical model method, flowing current method, simulated filter method, etc., have not been promoted and applied on a large scale due to limitations and unreliability of one kind or another.
目前各水厂还是根据进厂水流量，结合人工观察矾花效果和沉淀池的出口浊度仪来检验投药的效果，主要依靠流量比控制混凝剂投加量，无法做到水质的精确控制和检测。采用流量比的固定程序化自动控制***，试验结果只对取样瞬间水质有代表性，确定的混凝剂剂量存在不连续性及滞后性问题，很难在水厂运行过程中实现混凝剂投加量优化控制，无法实现混凝剂投加量的精确控 制。At present, each water plant still checks the effect of dosing according to the flow of water entering the plant, combined with manual observation of the effect of alum flowers and the outlet turbidimeter of the sedimentation tank. It mainly relies on the flow ratio to control the dosage of coagulant, and it is impossible to accurately control the water quality. and detection. A fixed programmed automatic control system with a flow ratio is used. The test results are only representative of the water quality at the moment of sampling. The determined coagulant dosage has problems of discontinuity and hysteresis. It is difficult to realize the coagulant dosage during the operation of the water plant. The optimal control of dosage cannot realize the precise control of coagulant dosage.
为了有效地形成一套更为客观的、综合各种因素的、将人的主观能动性和经验***化的解决方案。近年来混凝剂投加的智能化控制一直是水务工作者的关注重点，并且越来越倾向于智能控制方法与先进的混凝控制技术相结合的综合应用，以便实现混凝剂投加的最优控制。水下颗粒物摄像技术和计算机人工智能控制方法是实现混凝投药的智能化控制的一个较好的研究方向。随着人工智能、尤其是深度学***具有重要意义。In order to effectively form a set of more objective solutions that integrate various factors and systematize people's subjective initiative and experience. In recent years, the intelligent control of coagulant dosing has always been the focus of water workers, and more and more tends to the comprehensive application of the combination of intelligent control methods and advanced coagulation control technology in order to realize the coagulant dosing. best control. Underwater particle camera technology and computer artificial intelligence control method are a good research direction to realize the intelligent control of coagulation dosing. With the rise of artificial intelligence, especially deep learning, deep convolutional neural network models can be efficiently applied to image recognition, object detection and other fields. It can not only extract the texture features of the alum flower image, but also realize the purpose of recognizing the alum flower image. In addition, threshold segmentation and morphological processing can be performed on the collected alum flower images, and the main image features can be selected from many image features for standardization processing. Using the combination of SVM algorithm, BP neural network and GRNN neural network, the standardized features Conduct training to judge whether the amount of alum is appropriate. Computer vision realized through deep learning has been able to help people monitor the production environment, identify product defects, and identify potential failures. Therefore, this technology basically has the conditions for alum flower identification in the water production process of water plants, helping to achieve a more accurate image of alum flowers. Sophisticated and intelligent analysis. By using an active, automatic and self-learning system in the process of water production process to improve efficiency, reduce the loss of "electricity and medicine", achieve precise control of coagulation effect and real-time monitoring of water quality, which is very important for upgrading water plants Management level is of great significance.
基于机器学习和图像处理的传统矾花检测算法，主要包括数字图像处理模块、特征工程模块以及机器学习模块，如图1所示。在算法复杂程度上，由于 其算法逻辑串行且需要重复特征工程和机器学习直到符合相应场景，算法运行速度较慢且重复工作量大；在鲁棒性上，当场景发生变化，算法需重新设计，算法鲁棒性较差。The traditional alum flower detection algorithm based on machine learning and image processing mainly includes digital image processing module, feature engineering module and machine learning module, as shown in Figure 1. In terms of algorithm complexity, due to its serial algorithm logic and the need to repeat feature engineering and machine learning until it meets the corresponding scene, the algorithm runs slowly and has a large amount of repetitive work; in terms of robustness, when the scene changes, the algorithm needs to be re-started. Design, algorithm robustness is poor.
发明内容Contents of the invention
为克服上述现有技术存在的不足，本发明之目的在于提供一种利用图像识别技术分析矾花特征的方法及装置，以实现混凝剂投加的智能化控制，以便实现混凝剂的精确投加。In order to overcome the deficiencies in the above-mentioned prior art, the purpose of the present invention is to provide a method and device for analyzing the characteristics of alum flowers using image recognition technology, so as to realize the intelligent control of coagulant dosing, so as to realize the precise control of coagulant. add.
为达上述目的，本发明提出一种利用图像识别技术分析矾花特征的方法，包括如下步骤：For reaching above-mentioned purpose, the present invention proposes a kind of method utilizing image recognition technology to analyze alum flower feature, comprises the steps:
步骤S1，利用水下摄像头采集矾花图像，对采集图像打标，生成训练数据，建立矾花识别AI模型，利用打标数据作为监督信号，对矾花识别AI模型进行自主训练，获得训练好的矾花识别AI模型；Step S1, use the underwater camera to collect alum flower images, mark the collected images, generate training data, establish an AI model for alum flower recognition, use the marking data as a supervisory signal, conduct independent training on the alum flower recognition AI model, and obtain a well-trained Alum flower recognition AI model;
步骤S2，利用水下摄像头采集矾花图像，利用训练好的所述矾花识别AI模型进行识别；Step S2, using the underwater camera to collect images of alum flowers, and using the trained AI model for identifying alum flowers to identify them;
步骤S3，对单帧图像进行区域划分为M×M的区域，分别对单一区域计算矾花量化指标，并对单一区域各类矾花量化指标进行组合得到单帧图像的量化评价指标。In step S3, the single frame image is divided into M×M regions, and the quantitative index of alum flower is calculated for each single area, and the quantitative evaluation index of the single frame image is obtained by combining various quantitative indexes of alum flower in the single area.
优选地，步骤S1进一步包括：Preferably, step S1 further includes:
步骤S100，利用水下摄像头采集其景深范围内的矾花图像；Step S100, using the underwater camera to collect images of alum flowers within its depth of field;
步骤S101，确定各类别矾花标注标准，对采集图像打标，生成训练数据；Step S101, determining the labeling standards for each category of alum flowers, marking the collected images, and generating training data;
步骤S102，建立多视角矾花检测分割网络的矾花识别AI模型，利用打标数据作为监督信号，对多视角矾花检测分割网络进行训练，拟合标注数据。In step S102, an AI model for alum flower recognition of a multi-view alum flower detection and segmentation network is established, and the marking data is used as a supervisory signal to train the multi-view alum flower detection and segmentation network to fit the labeled data.
优选地，于步骤S101中，依据矾花的相关业务以及拍摄图像，确定三种类别矾花的标注标准，根据确定的标注标准对采集图像打标，对于在图像上无法确定蓬松矾花和片状矾花、矾花和模糊蓬松的图像，添加备注数据。Preferably, in step S101, according to the related business of alum flowers and the captured images, determine the labeling standards of the three types of alum flowers, and mark the collected images according to the determined labeling standards. Shaped alum flowers, alum flowers, and fuzzy fluffy images, add note data.
优选地，于步骤S102中，所述矾花识别AI模型采用多层神经网络，通过多层卷积与激活层，提取图像的高级特征，并利用所提取的图像高级特征，通过回归任务回归得出矾花的检测框，以及利用图像高级特征，通过分类任务分类得出矾花的类型。Preferably, in step S102, the alum flower recognition AI model uses a multi-layer neural network to extract high-level features of the image through multiple layers of convolution and activation layers, and use the extracted high-level features of the image to obtain Get the detection frame of the alum flower, and use the advanced features of the image to classify the type of alum flower through the classification task.
优选地，于步骤S102中，通过大量矾花数据，利用反向传播优化模型参数，训练所述矾花识别AI模型。Preferably, in step S102, a large amount of alum flower data is used to optimize model parameters through back propagation, and to train the alum flower recognition AI model.
优选地，于步骤S2中，利用所述训练好的矾花识别AI模型识别采集图像，检出每个矾花类别、概率以及位置等属性。Preferably, in step S2, the trained alum flower recognition AI model is used to identify and collect images, and attributes such as category, probability and location of each alum flower are detected.
优选地，所述单一区域每一矾花类别的量化指标包括该类别在该区域的个数、置信平均、置信中位、面积平均以及面积中位。Preferably, the quantitative index of each alum flower category in the single region includes the number of the category in the region, confidence average, confidence median, area average and area median.
优选地，于步骤S3中，将M×M的所有区域三类矾花量化指标进行组合得到15M 2维的矾花量化指标，即为单帧图像的量化评价指标。 Preferably, in step S3, the quantitative indicators of three types of alum flowers in all areas of M×M are combined to obtain 15M 2 -dimensional quantitative indicators of alum flowers, which are the quantitative evaluation indicators of a single frame image.
优选地，于步骤S3中，从1分钟内采集的图像等间隔选取F帧，计算单帧 图像的矾花量化指标，然后将F帧相同矾花量化指标求平均，得到时间量化后的矾花量化指标。Preferably, in step S3, F frames are selected at equal intervals from the images collected within 1 minute, and the quantitative index of alum flower of a single frame image is calculated, and then the same quantitative index of alum flower in F frames is averaged to obtain the time-quantified alum flower Quantitative indicators.
为达到上述目的，本发明还提供一种利用图像识别技术分析矾花特征的装置，包括：In order to achieve the above object, the present invention also provides a device utilizing image recognition technology to analyze the characteristics of alum flowers, comprising:
矾花识别AI模型构建及训练单元，用于利用水下摄像头采集矾花图像，对采集图像打标，生成训练数据，建立矾花识别AI模型，利用打标数据作为监督信号，对矾花识别AI模型进行自主训练，获得训练好的矾花识别AI模型；Alum flower recognition AI model construction and training unit, used to use underwater camera to collect alum flower images, mark the collected images, generate training data, establish alum flower recognition AI model, use the marking data as a supervisory signal, and identify alum flower The AI model is trained independently to obtain the trained AI model for alum flower recognition;
图像采集识别单元，用于利用水下摄像头采集矾花图像，利用训练好的所述矾花识别AI模型进行识别；The image collection and recognition unit is used to collect the alum flower image by using the underwater camera, and use the trained AI model to identify the alum flower;
量化评价指标检测单元，用于对单帧图像进行区域划分为M×M的区域，分别对单一区域计算矾花量化指标，并对单一区域三类矾花量化指标进行组合得到单帧图像的量化评价指标。The quantitative evaluation index detection unit is used to divide a single frame image into M×M areas, calculate the quantitative index of alum flower for a single area respectively, and combine the three types of alum flower quantitative indicators in a single area to obtain the quantification of a single frame image evaluation index.
优选地，所述量化评价指标检测单元从1分钟内采集的图像等间隔选取F帧，计算单帧图像的矾花量化指标，然后将F帧相同矾花量化指标求平均，得到时间量化后的矾花量化指标。Preferably, the quantitative evaluation index detection unit selects F frames at equal intervals from the images collected within 1 minute, calculates the quantitative index of alum flower in a single frame image, and then averages the same quantitative index of alum flower in F frames to obtain the time-quantified Quantitative indicators of alum flowers.
与现有技术相比，本发明一种利用图像识别技术分析矾花特征的方法及装置，通过构建矾花识别AI模型实现对矾花图像进行识别，并检出矾花量化指标，从而可以判断出投矾量是否合适，以实现混凝剂投加的智能化控制，以便实现混凝剂的精确投加。Compared with the prior art, the present invention is a method and device for analyzing the characteristics of alum flowers using image recognition technology. By constructing an AI model for alum flower recognition, the image of alum flowers can be recognized, and the quantitative indicators of alum flowers can be detected, so that it can be judged Whether the amount of alum is appropriate to realize the intelligent control of coagulant dosing, so as to realize the precise dosing of coagulant.
附图说明Description of drawings
图1为基于机器学习和图像处理的传统矾花检测算法的示意图；Fig. 1 is the schematic diagram of the traditional alum flower detection algorithm based on machine learning and image processing;
图2a为本发明中感知器(神经元)的结构图；Fig. 2 a is the structural diagram of perceptron (neuron) among the present invention;
图2b为本发明中多层感知器的结构图；Figure 2b is a structural diagram of a multi-layer perceptron in the present invention;
图2c为本发明中多层感知器的学习(训练)过程示意图；Fig. 2c is a schematic diagram of the learning (training) process of the multi-layer perceptron in the present invention;
图3为本发明一种利用图像识别技术分析矾花特征的方法的步骤流程图；Fig. 3 is a flow chart of the steps of a method utilizing image recognition technology to analyze the characteristics of alum flowers of the present invention;
图4为原水净化过程示意图；Fig. 4 is the schematic diagram of raw water purification process;
图5为本发明实施例中片状矾花、蓬松矾花以及模糊蓬松示意图；Fig. 5 is a schematic diagram of flaky alum flowers, fluffy alum flowers and fuzzy fluffy in the embodiment of the present invention;
图6为本发明实施例中简单光学***景深说明示意图；Fig. 6 is a schematic diagram illustrating the depth of field of a simple optical system in an embodiment of the present invention;
图7为本发明实施例中三类矾花标注标准示意图；Fig. 7 is a standard schematic diagram of three types of alum flowers in the embodiment of the present invention;
图8为本发明具体实施例中矾花识别AI模型采用的多层神经网络的结构示意图；Fig. 8 is the structural representation of the multilayer neural network that alum flower recognition AI model adopts in the specific embodiment of the present invention;
图9为本发明实施例中多视角矾花检测分割网络示意图；9 is a schematic diagram of a multi-view alum flower detection and segmentation network in an embodiment of the present invention;
图10为本发明实施例中单帧图像2×2方式划分图像示意图；FIG. 10 is a schematic diagram of dividing an image in a 2×2 manner for a single frame image in an embodiment of the present invention;
图11为本发明实施例中单一区域矾花量化指标计算示意图；Fig. 11 is a schematic diagram of calculation of quantitative index of alum flower in a single region in the embodiment of the present invention;
图12为本发明一种利用图像识别技术分析矾花特征的装置的***架构图；12 is a system architecture diagram of a device for analyzing the characteristics of alum flowers using image recognition technology according to the present invention;
图13为本发明实施例中矾花量化各个指标对于沉后水浊度的重要程度示意图；Fig. 13 is a schematic diagram of the importance of various indicators of alum flower quantification for sinking water turbidity in the embodiment of the present invention;
图14为本发明实施例中所有区域不同类别矾花重要性示意图。Figure 14 is a schematic diagram of the importance of different types of alum flowers in all regions in the embodiment of the present invention.
具体实施方式Detailed ways
以下通过特定的具体实例并结合附图说明本发明的实施方式，本领域技术人员可由本说明书所揭示的内容轻易地了解本发明的其它优点与功效。本发明亦可通过其它不同的具体实例加以施行或应用，本说明书中的各项细节亦可基于不同观点与应用，在不背离本发明的精神下进行各种修饰与变更。The implementation of the present invention is described below through specific examples and in conjunction with the accompanying drawings, and those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific examples, and various modifications and changes can be made to the details in this specification based on different viewpoints and applications without departing from the spirit of the present invention.
在介绍本发明之前，先介绍一下卷积神经络基础原理：Before introducing the present invention, first introduce the basic principles of convolutional neural network:
感知器(神经元)，其结构如图2a所示，具体如下：Perceptron (neuron), its structure is as shown in Figure 2a, specifically as follows:
·感知器包含一组参数，包含线性计算和非线性计算；The perceptron contains a set of parameters, including linear calculations and nonlinear calculations;
·感知器是具有单一人造神经元的神经网络，它有一个输入层，和将输入单元和输出单元相连的一组连接；A perceptron is a neural network with a single artificial neuron, an input layer, and a set of connections connecting the input unit to the output unit;
·感知器的目标是对提供给输入单元的图案进行分类。• The goal of the perceptron is to classify the pattern presented to the input unit.
·输出单元执行的基本操作是，把每个输入(xn)与其连接强度或权重(wn)相乘，并将乘积的总和传递给输出单元；The basic operation performed by the output unit is to multiply each input (xn) with its connection strength or weight (wn) and pass the sum of the products to the output unit;
·输入的加权和
Figure PCTCN2021119244-appb-000001
与阈值θ进行比较后的结果被传递给阶跃函数(激活函数)。如果总和超过阈值，则阶跃函数输出“1”，否则输出“0”。 · Weighted sum of inputs
Figure PCTCN2021119244-appb-000001
The result of comparison with the threshold θ is passed to the step function (activation function). The step function outputs "1" if the sum exceeds the threshold, and "0" otherwise.
例如，输入可以是图像中像素，或者更常见的情况是，从原始图像中提取的特征，例如图像中对象的轮廓。每次输入一个图像，感知器会判定该图像是否为某类别的成员，例如猫类。输出只能是两种状态之一，如果图像处于类别中，则为“开”，否则为“关”。“开”和“关”分别对应二进制值中的1和0。For example, the input can be pixels in an image, or more commonly, features extracted from the original image, such as the contours of objects in the image. Each time an image is input, the perceptron determines whether the image is a member of a class, such as cats. The output can only be one of two states, 'on' if the image is in a category, 'off' otherwise. "On" and "Off" correspond to 1 and 0 in binary values, respectively.
多层感知器，其结构如图2b所示，具体如下：Multi-layer perceptron, its structure is shown in Figure 2b, specifically as follows:
包含多个层次的感知器的前馈网络，前馈网络的逐层计算：Feedforward network containing multiple layers of perceptrons, layer-by-layer calculation of feedforward network:
输入值从输入层神经元通过加权连接逐层前向传播，经过隐含层，最后到达输出层得到输出。在信号的前向传播过程中，网络的权值是固定不变的，每一层神经元的状态只影响下一层神经元的状态。其过程如下公式：The input value propagates forward layer by layer from the input layer neurons through the weighted connection, passes through the hidden layer, and finally reaches the output layer to get the output. During the forward propagation of the signal, the weight of the network is fixed, and the state of neurons in each layer only affects the state of neurons in the next layer. The process is as follows:
Figure PCTCN2021119244-appb-000002
Figure PCTCN2021119244-appb-000002
Figure PCTCN2021119244-appb-000003
Figure PCTCN2021119244-appb-000003
Figure PCTCN2021119244-appb-000004
Figure PCTCN2021119244-appb-000004
Figure PCTCN2021119244-appb-000005
Figure PCTCN2021119244-appb-000005
多层感知器的学习(训练)过程如图2c所示，具体如下：The learning (training) process of the multi-layer perceptron is shown in Figure 2c, as follows:
·前向传播为通过感知器计算输出；Forward propagation is to calculate the output through the perceptron;
·反向传播为训练感知器参数的过程，反向传播的输入是被前向传播的前馈；Backpropagation is the process of training perceptron parameters, and the input of backpropagation is the feedforward of forward propagation;
·输出结果与训练者给出的值进行比较，差值被用来更新连接输出单元的权重，以减少误差；The output result is compared with the value given by the trainer, and the difference is used to update the weight of the connected output unit to reduce the error;
·根据每个权重对误差贡献的多少，通过反向传播误差对输入单元和隐藏层之间的权重进行更新。The weights between the input unit and the hidden layer are updated by backpropagating the error according to how much each weight contributes to the error.
·利用大量样本进行训练，隐藏单元生成了可区分不同输入模式的选择性特征，这样一来它们就能够在输出层中对不同类别进行区分。Training with a large number of examples, the hidden units generate selective features that distinguish different input patterns, so that they can distinguish between different categories in the output layer.
图3为本发明一种利用图像识别技术分析矾花特征的方法的步骤流程图。 如图3所示，本发明一种利用图像识别技术分析矾花特征的方法，包括如下步骤：Fig. 3 is a flow chart of steps of a method for analyzing the characteristics of alum flowers using image recognition technology in the present invention. As shown in Figure 3, a kind of method of utilizing image recognition technology to analyze alum flower feature of the present invention comprises the following steps:
步骤S1，利用水下摄像头采集矾花图像，对采集图像打标，生成训练数据，建立矾花识别AI模型，利用打标数据作为监督信号，对矾花识别AI模型进行自主训练，从而获得训练好的矾花识别AI模型。Step S1, use the underwater camera to collect alum flower images, mark the collected images, generate training data, establish an AI model for alum flower recognition, and use the marking data as a supervisory signal to independently train the alum flower recognition AI model to obtain training data. A good alum flower recognition AI model.
一般地，原水的净化主要是在絮凝池、沉淀池以及滤池进行，如图4所示，在絮凝池加入混凝剂通过化学反应生成颗粒状的晶体矾花，水流通过絮凝池进入沉淀池，在沉淀池入水口处矾花基本上生成完成，颗粒状矾花在混凝池和沉淀池中，不断地相互结合、运动以及吸附原水杂质，其形态由颗粒状到片状，片状矾花内部相互运动吸附形成蓬松矾花，此时矾花密度大于水的密度，最后由于重力作用沉淀到沉淀池底部，完成初步净水，降低原水浊度。通过矾花业务分析，根据其形态，区分为片状矾花和蓬松矾花。但是由于摄像头焦距等参数固定，靠近摄像头的矾花会在图像上呈现模糊状，导致人眼无法区分其具体形态，对于这部分矾花，不能简单的归类为片状和蓬松矾花，而是需要单独归类为模糊蓬松，作为矾花的补充类别。因此，考虑到矾花的业务以及摄像头的影响，在矾花识别AI模型中，将矾花分为：片状矾花、蓬松矾花以及模糊蓬松，如图5所示。Generally, the purification of raw water is mainly carried out in flocculation tanks, sedimentation tanks and filter tanks. As shown in Figure 4, coagulants are added to the flocculation tanks to form granular crystal alum flowers through chemical reactions, and the water flows through the flocculation tanks into the sedimentation tanks. , the alum flowers are basically formed at the water inlet of the sedimentation tank. The granular alum flowers are constantly combined, moving and adsorbing impurities in the raw water in the coagulation tank and the sedimentation tank. The flowers move and absorb each other to form fluffy alum flowers. At this time, the density of alum flowers is greater than that of water, and finally they settle to the bottom of the sedimentation tank due to gravity, completing preliminary water purification and reducing the turbidity of raw water. Through the business analysis of alum flowers, according to their shape, they can be divided into flaky alum flowers and fluffy alum flowers. However, due to the fixed parameters such as the focal length of the camera, the alum flowers close to the camera will appear blurred on the image, causing the human eye to be unable to distinguish their specific shapes. For this part of the alum flowers, they cannot be simply classified as flaky and fluffy alum flowers, but It needs to be classified separately as fuzzy and fluffy, as a supplementary category of alum flower. Therefore, considering the business of alum flowers and the influence of cameras, in the AI model of alum flower recognition, alum flowers are divided into: flaky alum flowers, fluffy alum flowers, and fuzzy and fluffy, as shown in Figure 5.
具体地，步骤S1进一步包括：Specifically, step S1 further includes:
步骤S100，利用水下摄像头采集其景深范围内的矾花图像。Step S100, using the underwater camera to collect images of alum flowers within its depth of field.
以图6的简单光学***为例介绍水下摄像头的拍摄的立体范围，理论上，正焦距的光学***可以实现无限远成像，但是由于光路能量损耗以及CCD(光学传感器)相元尺寸原因无法做到相机无限远成像。规定相机支持最小弥散斑直径为δ，通过摄像头的焦距、相元尺寸以及光学***相关参数可计算出，摄像头光学***的景深。景深其物理意义：物体在物平面景深范围内能够在像平面(CCD)能够得到相对清晰的像。因此，水下摄像头拍摄范围为景深附近的长方体内的矾花，在本发明具体实施例中，水下摄像头采用两个摄像头，放置在沉淀池入水口且安装在三管托板上，距离水面80cm，斜向下45度角拍摄。Take the simple optical system in Figure 6 as an example to introduce the three-dimensional range of the underwater camera. In theory, the optical system with positive focal length can realize infinite distance imaging, but it cannot be done due to the energy loss of the optical path and the size of the CCD (optical sensor) phase element. Imaging at infinity to the camera. It is stipulated that the minimum diameter of the diffuse spot supported by the camera is δ, and the depth of field of the optical system of the camera can be calculated through the focal length of the camera, the size of the phase element, and the relevant parameters of the optical system. The physical meaning of the depth of field: the object can get a relatively clear image on the image plane (CCD) within the depth of field of the object plane. Therefore, the shooting range of the underwater camera is the alum flowers in the cuboid near the depth of field. In a specific embodiment of the present invention, the underwater camera adopts two cameras, which are placed at the water inlet of the sedimentation tank and installed on the three-tube supporting plate, with a distance from the water surface 80cm, shot at an angle of 45 degrees obliquely downward.
步骤S101，确定各类别矾花标注标准，对采集图像打标，生成训练数据。Step S101, determining the labeling standards for each category of alum flowers, marking the collected images, and generating training data.
在建立矾花识别AI模型之前，需要有一定量的打标数据作为监督信号，并让矾花识别AI模型自主训练，拟合标注数据，因此，标注数据是矾花识别AI模型比较重要的环节。结合矾花业务以及拍摄情况，将矾花更加细粒度的区分。在标注上，引入了一个问题：如何根据拍摄图像区分各类别矾花。Before establishing the alum flower recognition AI model, a certain amount of marking data is required as a supervisory signal, and the alum flower recognition AI model is allowed to train independently and fit the labeled data. Therefore, labeling data is an important part of the alum flower recognition AI model. Combined with the alum flower business and the shooting situation, the alum flower is more fine-grained. In terms of labeling, a problem is introduced: how to distinguish various types of alum flowers according to the captured images.
在实际情况中，片状与蓬松矾花均为人为定义，在之前的研究中并没有对矾花类别做细粒度区分，两种矾花间在行业中没有严格明确的定义，在图像视觉上也没有明确的分界线，同时在以往的工作中，只是通过人眼在沉淀池边上观察矾花。因此，在本发明中，依据矾花的相关业务以及拍摄图像，确定了一套相对完整的3种类别矾花的标注标准，如图7所示，从而根据确定的标注标准对采集图像打标，生成训练数据。In the actual situation, flake and fluffy alum flowers are both artificially defined. In the previous research, there was no fine-grained distinction between the alum flower categories. There is no strict and clear definition between the two kinds of alum flowers in the industry. There is no clear dividing line, and at the same time, in the previous work, the alum flowers were only observed by human eyes on the side of the sedimentation tank. Therefore, in the present invention, based on the related business of alum flowers and the captured images, a relatively complete set of labeling standards for three types of alum flowers is determined, as shown in Figure 7, so that the collected images are marked according to the determined labeling standards , to generate training data.
从图7中，可以看到在标注标准中，存在备注记录，代表在图像上无法确定蓬松矾花和片状矾花、矾花和模糊蓬松。对于备注数据，在数据预处理之前，根据其标签记录，给出相应标签的概率值，表示其类别不确定性。其目的在于：(1)符合人类认知，引导矾花识别AI模型学习人类对矾花的普适性认知；(2)在数据层面引入矾花类别不确定性，增强模型鲁棒性。From Figure 7, it can be seen that in the labeling standard, there is a note record, which means that fluffy alum flowers, flaky alum flowers, alum flowers and fuzzy fluffy cannot be determined on the image. For the remark data, before data preprocessing, according to its label record, the probability value of the corresponding label is given, indicating its category uncertainty. Its purpose is: (1) conform to human cognition, guide the AI model of alum flower recognition to learn the universal cognition of human beings to alum flower; (2) introduce the uncertainty of alum flower category at the data level, and enhance the robustness of the model.
步骤S102，建立多视角矾花检测分割网络的矾花识别AI模型，利用打标数据作为监督信号，对多视角矾花检测分割网络进行训练，拟合标注数据。In step S102, an AI model for alum flower recognition of a multi-view alum flower detection and segmentation network is established, and the marking data is used as a supervisory signal to train the multi-view alum flower detection and segmentation network to fit the labeled data.
在本发明具体实施例中，所述矾花识别AI模型采用多层神经网络，如图8所示，在该多层神经网络中，卷积层+激活函数可看作感知器，整个网络包含多层感知器，特征经过多个卷积层+激活函数(感知器)，不断进行提取和压缩的，最终能得到比较高层次特征，即通过该多层神经网络对原始特征一步又一步的浓缩，最终得到的特征更可靠，最后利用最后一层特征可以做各种任务：比如分类(如判断类别)、回归(如回归出检测框)等。In a specific embodiment of the present invention, the alum flower recognition AI model adopts a multilayer neural network, as shown in Figure 8, in this multilayer neural network, the convolution layer+activation function can be regarded as a perceptron, and the whole network includes Multi-layer perceptron, features are continuously extracted and compressed through multiple convolutional layers + activation functions (perceptrons), and finally relatively high-level features can be obtained, that is, the original features are concentrated step by step through the multi-layer neural network , the final features are more reliable, and finally use the last layer of features to do various tasks: such as classification (such as judging the category), regression (such as returning the detection frame), etc.
在本发明中，所述矾花识别AI模型的工作原理主要是：In the present invention, the operating principle of the alum flower recognition AI model is mainly:
1、通过多层卷积+激活层，提取图像的高级特征；1. Extract the advanced features of the image through multi-layer convolution + activation layer;
2、利用图像高级特征，通过回归任务，回归得出矾花的检测框(位置和大小)；2. Using the advanced features of the image, through the regression task, the detection frame (position and size) of the alum flower is obtained by regression;
3、利用图像高级特征，通过分类任务，分类得出矾花的类型(蓬松、片状、模糊矾花)。3. Using the advanced features of the image, through the classification task, the type of alum flower (fluffy, flaky, fuzzy alum flower) can be obtained by classification.
其训练过程如下：Its training process is as follows:
·收集大量矾花图像，并进进行人工打标(标注矾花位置、大小、类型)Collect a large number of alum flower images and carry out manual marking (marking the position, size and type of alum flowers)
·通过大量矾花数据，利用反向传播优化模型参数，训练矾花模型(多层神经网络)。·Through a large amount of alum flower data, use backpropagation to optimize the model parameters and train the alum flower model (multi-layer neural network).
相对于传统的机器学***台，将实战经验和学术引入算法中；在视频信息方面，本发明通过光流法、视频时空特征提取技术，能够提升视频信息提取和集成，提升检测分割算法的精度和鲁棒性。如图9所示，在本发明通过多视角基于深度学习的矾花检测分割算法，不同水下拍摄条件下都能够对矾花有比较好的检出。Compared with the traditional machine learning algorithm, the alum flower detection and segmentation algorithm based on deep learning can automatically extract and combine features without manual adjustment of the algorithm. The present invention precipitates detection and segmentation algorithms and industrial vision platforms in actual combat, and introduces practical experience and academic knowledge into the algorithm; in terms of video information, the present invention can improve video information extraction and integration through optical flow method and video spatiotemporal feature extraction technology, Improve the accuracy and robustness of detection and segmentation algorithms. As shown in FIG. 9 , in the present invention, the alum flower detection and segmentation algorithm based on multi-view and deep learning can detect alum flower better under different underwater shooting conditions.
步骤S2，利用水下摄像头采集矾花图像，利用训练好的所述矾花识别AI模型进行识别。Step S2, using the underwater camera to collect images of alum flowers, and using the trained AI model for identifying alum flowers to identify them.
在本发明中，利用所述训练好的矾花识别AI模型能够检测拍摄图像中每个检出矾花类别、概率以及位置等属性。检出属性列表如下：In the present invention, the trained alum flower recognition AI model can detect attributes such as category, probability and position of each detected alum flower in the captured image. The list of checked out properties is as follows:
表1检测分割网络检出矾花相关属性表Table 1 The related attributes of alum flowers detected by the detection and segmentation network
Figure PCTCN2021119244-appb-000006
Figure PCTCN2021119244-appb-000006
步骤S3，对单帧图像进行区域划分为M×M的区域，分别对单一区域计算矾花量化指标，并对单一区域3类矾花量化指标进行组合得到单帧图像的量化评价指标，其中单一区域每一矾花类别的量化指标包括该类别在该区域的个数、 置信平均、置信中位、面积平均以及面积中位。Step S3, divide the single-frame image into M×M regions, calculate the quantitative index of the alum flower for the single area, and combine the three types of alum flower quantitative indicators in the single area to obtain the quantitative evaluation index of the single-frame image, in which the single The quantitative indicators of each alum flower category in the region include the number of the category in the region, the confidence average, the confidence median, the area average, and the area median.
由于拍摄图像区域内矾花分布不均匀且为了保留矾花分布信息，在量化矾花特征之前，对单帧图像进行区域划分为M×M的区域。以2×2为例，如图10所示，将整张图像以2×2的形式划分成如图所示的4个区域，然后分别对单一区域计算矾花量化指标。Due to the uneven distribution of alum flowers in the captured image area and in order to preserve the distribution information of alum flowers, before quantifying the characteristics of alum flowers, the single frame image was divided into M×M regions. Taking 2×2 as an example, as shown in Figure 10, the entire image is divided into 4 regions as shown in the figure in the form of 2×2, and then the quantitative index of alum flower is calculated for a single region.
单帧单一区域矾花量化指标计算：以图10左上区域为例来说明单帧单一区域矾花量化指标计算方式。如图11所示，对单一区域内，每一类矾花类别分别计算该类别在该区域的个数、置信平均、置信中位、面积平均以及面积中位。Calculation of quantitative indicators of alum flower in single frame and single area: take the upper left area of Figure 10 as an example to illustrate the calculation method of quantitative index of alum flower in single frame and single area. As shown in Figure 11, for each type of alum flower category in a single area, the number, confidence average, confidence median, area average, and area median of the category in the area are calculated respectively.
具体地，置信平均数、置信中位数计算方式：Specifically, the calculation method of confidence mean and confidence median:
Figure PCTCN2021119244-appb-000007
Figure PCTCN2021119244-appb-000007
其中，
Figure PCTCN2021119244-appb-000008
表示c类矾花的平均置信，
Figure PCTCN2021119244-appb-000009
表示c类矾花的置信中位数， Nc表示该区域内c类矾花的个数。 in,
Figure PCTCN2021119244-appb-000008
Indicates the average confidence of class c alum flower,
Figure PCTCN2021119244-appb-000009
Indicates the confidence median of type c alum flowers, and N c indicates the number of c type alum flowers in this area.
面积平均数、面积中位数计算方式：Calculation method of area mean and area median:
Figure PCTCN2021119244-appb-000010
Figure PCTCN2021119244-appb-000010
其中，
Figure PCTCN2021119244-appb-000011
表示c类矾花的平均面积，
Figure PCTCN2021119244-appb-000012
表示c类矾花的面积中位数，A表示该区域的面积。 in,
Figure PCTCN2021119244-appb-000011
Indicates the average area of c-type alum flowers,
Figure PCTCN2021119244-appb-000012
Indicates the median area of type c alum flowers, and A indicates the area of this area.
因此，单一区域每一矾花类别的量化指标：Therefore, the quantitative indicators for each type of alum in a single region:
Figure PCTCN2021119244-appb-000013
Figure PCTCN2021119244-appb-000013
其中，c代表矾花类别，i代表划分区域。Among them, c represents the category of alum flowers, and i represents the division area.
在本发明中，矾花各指标代表的物理意义：In the present invention, the physical significance that each index of alum flower represents:
(1)个数：当前区域某类检出矾花的个数；(1) Number: the number of a certain type of alum flowers detected in the current area;
(2)置信平均、置信中位：置信即矾花识别AI模型预测检出矾花的概率，表征检出矾花为当前类别的置信度；(2) Confidence average, confidence median: Confidence is the probability that the alum flower recognition AI model predicts the detection of alum flower, which represents the confidence that the detected alum flower is the current category;
(3)面积平均、面积中位：当前区域某类检出矾花面积之和相对于该区域的面积的比值，代表该区域该类矾花的密集程度。(3) Average area and median area: The ratio of the sum of the area of a certain type of alum flowers detected in the current area to the area of the area represents the density of this type of alum flowers in the area.
由上可知，单一区域每类矾花具有5个量化评价指标，根据矾花类别划分可得单一区域矾花指标共计5×3＝15个。It can be seen from the above that each type of alum flower in a single region has 5 quantitative evaluation indicators, and according to the classification of alum flower categories, a total of 5 × 3 = 15 alum flower indicators in a single region can be obtained.
单帧多区域矾花量化指标组合：将M×M的所有区域3类矾花量化指标进行组合可得到15M 2维的矾花量化指标，即为单帧图像的量化评价指标。 Combination of single-frame and multi-region alum quantitative indicators: Combining the three types of alum quantitative indicators in all regions of M×M can obtain 15M 2- dimensional alum quantitative indicators, which are quantitative evaluation indicators for single-frame images.
Figure PCTCN2021119244-appb-000014
Figure PCTCN2021119244-appb-000014
以图10的2×2划分为例，所提取的矾花量化指标个数为15×2×2＝60个。Taking the 2×2 division in Figure 10 as an example, the number of quantitative indicators of alum flowers extracted is 15×2×2=60.
矾花量化指标时间量化：由于原水数据1分钟更新一次，而由于摄像头FPS特征，1分钟约有360帧图像，若所有图像都参与矾花量化指标的计算，则会导致：(1)每帧图像都要经过矾花识别AI模型以及单帧多区域矾花量化评价的计算，计算复杂度过大；(2)相邻帧矾花量化指标相近，特征冗余。Time quantification of alum quantitative indicators: Since the raw water data is updated once a minute, and due to the FPS characteristics of the camera, there are about 360 frames of images in one minute. If all images participate in the calculation of alum quantitative indicators, it will result in: (1) every frame The images have to go through the calculation of the AI model of alum flower recognition and the quantitative evaluation of multi-region alum flower in a single frame, and the calculation complexity is too large; (2) the quantitative indicators of alum flower in adjacent frames are similar, and the features are redundant.
因此，在实际计算1分钟的矾花量化指标时，需要对时间量化，采用的方式：从1分钟内采集的图像等间隔选取F帧，计算单帧图像的矾花量化指标， 最后F帧相同矾花量化指标求平均，即可得到时间量化后的矾花量化指标。Therefore, when actually calculating the quantitative index of alum flower for 1 minute, it is necessary to quantify the time. The method adopted is to select F frames at equal intervals from the images collected within 1 minute, and calculate the quantitative index of alum flower for a single frame image. Finally, the F frames are the same The quantitative index of the alum flower is averaged to obtain the quantitative index of the alum flower after time quantification.
图12为本发明一种利用图像识别技术分析矾花特征的装置的***架构图。如图12所示，本发明一种利用图像识别技术分析矾花特征的装置，包括：Fig. 12 is a system architecture diagram of a device for analyzing characteristics of alum flowers using image recognition technology according to the present invention. As shown in Figure 12, a kind of device of the present invention utilizes image recognition technology to analyze the characteristics of alum flowers, comprising:
矾花识别AI模型构建及训练单元10，用于利用水下摄像头采集矾花图像，对采集图像打标，生成训练数据，建立矾花识别AI模型，利用打标数据作为监督信号，对矾花识别AI模型进行自主训练，获得训练好的矾花识别AI模型。The alum flower recognition AI model construction and training unit 10 is used to collect alum flower images with an underwater camera, mark the collected images, generate training data, establish an alum flower recognition AI model, use the marking data as a supervisory signal, and identify the alum flower Recognize the AI model for autonomous training, and obtain the trained AI model for alum flower recognition.
一般地，原水的净化主要是在絮凝池、沉淀池以及滤池进行，在絮凝池加入混凝剂通过化学反应生成颗粒状的晶体矾花，水流通过絮凝池进入沉淀池，在沉淀池入水口处矾花基本上生成完成，颗粒状矾花在混凝池和沉淀池中，不断地相互结合、运动以及吸附原水杂质，其形态由颗粒状到片状，片状矾花内部相互运动吸附形成蓬松矾花，此时矾花密度大于水的密度，最后由于重力作用沉淀到沉淀池底部，完成初步净水，降低原水浊度。通过矾花业务分析，根据其形态，区分为片状矾花和蓬松矾花。但是由于摄像头焦距等参数固定，靠近摄像头的矾花会在图像上呈现模糊状，导致人眼无法区分其具体形态，对于这部分矾花，不能简单的归类为片状和蓬松矾花，而是需要单独归类为模糊蓬松，作为矾花的补充类别。因此，考虑到矾花的业务以及摄像头的影响，在矾花识别AI模型中，将矾花分为：片状矾花、蓬松矾花以及模糊蓬松，如图5所示。Generally, the purification of raw water is mainly carried out in flocculation tanks, sedimentation tanks and filter tanks. Coagulants are added to the flocculation tanks to form granular crystal alum flowers through chemical reactions. The water flows into the sedimentation tanks through the flocculation tanks. The formation of alum flowers is basically completed. In the coagulation tank and the sedimentation tank, the granular alum flowers are constantly combined with each other, moving and adsorbing impurities in the raw water. The shape changes from granular to flake. Fluffy alum flowers, at this time the density of alum flowers is greater than that of water, and finally settle to the bottom of the sedimentation tank due to gravity, completing preliminary water purification and reducing the turbidity of raw water. Through the business analysis of alum flowers, according to their shape, they can be divided into flaky alum flowers and fluffy alum flowers. However, due to the fixed parameters such as the focal length of the camera, the alum flowers close to the camera will appear blurred on the image, causing the human eye to be unable to distinguish their specific shapes. For this part of the alum flowers, they cannot be simply classified as flaky and fluffy alum flowers, but It needs to be classified separately as fuzzy and fluffy, as a supplementary category of alum flower. Therefore, considering the business of alum flowers and the influence of cameras, in the AI model of alum flower recognition, alum flowers are divided into: flaky alum flowers, fluffy alum flowers, and fuzzy and fluffy, as shown in Figure 5.
具体地，矾花识别AI模型构建及训练单元10进一步包括：Specifically, alum flower recognition AI model construction and training unit 10 further include:
图像采集模块，用于利用水下摄像头采集其景深范围内的矾花图像。The image acquisition module is used to utilize the underwater camera to collect images of alum flowers within the depth of field.
在本发明中，水下摄像头拍摄范围为景深附近的长方体内的矾花。In the present invention, the shooting range of the underwater camera is the alum flowers in the cuboid near the depth of field.
打标模块，用于确定各类别矾花标注标准，对采集图像打标，生成训练数据。The marking module is used to determine the labeling standards of various types of alum flowers, mark the collected images, and generate training data.
在建立矾花识别AI模型之前，需要有一定量的打标数据作为监督信号，并让矾花识别AI模型自主训练，拟合标注数据，因此，标注数据是矾花识别AI模型比较重要的环节。结合矾花业务以及拍摄情况，将矾花更加细粒度的区分。在标注上，引入了一个问题：如何根据拍摄图像区分各类别矾花Before establishing the alum flower recognition AI model, a certain amount of marking data is required as a supervisory signal, and the alum flower recognition AI model is allowed to train independently and fit the labeled data. Therefore, labeling data is an important part of the alum flower recognition AI model. Combined with the alum flower business and the shooting situation, the alum flower is more fine-grained. In terms of labeling, a problem is introduced: how to distinguish various types of alum flowers according to the captured images
在实际情况中，片状与蓬松矾花均为人为定义，在之前的研究中并没有对矾花类别做细粒度区分，两种矾花间在行业中没有严格明确的定义，在图像视觉上也没有明确的分界线，同时水厂专业的老师在以往的工作中，只是通过人眼在沉淀池边上观察矾花。因此，在本发明中，依据矾花的相关业务以及拍摄图像，最终确定一套相对完整的3种类别矾花的标注标准，从而根据确定的标注标准对采集图像打标，生成训练数据。In the actual situation, flake and fluffy alum flowers are both artificially defined. In the previous research, there was no fine-grained distinction between the alum flower categories. There is no strict and clear definition between the two kinds of alum flowers in the industry. There is also no clear dividing line. At the same time, in the past work, the teachers of the water plant only observed the alum flowers by the side of the sedimentation tank with human eyes. Therefore, in the present invention, based on the related business of alum flowers and the captured images, a relatively complete set of labeling standards for three types of alum flowers is finally determined, so that the collected images are marked according to the determined labeling standards to generate training data.
训练模块，用于建立多视角矾花检测分割网络的矾花识别AI模型，利用打标数据作为监督信号，对多视角矾花检测分割网络进行训练，拟合标注数据。The training module is used to establish the alum flower recognition AI model of the multi-view alum flower detection and segmentation network, use the marking data as the supervision signal, train the multi-view alum flower detection and segmentation network, and fit the labeled data.
训练模块利用多层神经网络建立矾花识别AI模型，该多层神经网络如图8所示，在该多层神经网络中，卷积层+激活函数可看作感知器，整个网络包含多层感知器，特征经过多个卷积层+激活函数(感知器)，不断进行提取和压缩的，最终能得到比较高层次特征，即通过该多层神经网络对原始特征一步又一步的浓缩，最终得到的特征更可靠，最后利用最后一层特征可以做各种任务：比如 分类(如判断类别)、回归(如回归出检测框)等。The training module uses a multi-layer neural network to establish an AI model for alum flower recognition. The multi-layer neural network is shown in Figure 8. In this multi-layer neural network, the convolution layer + activation function can be regarded as a perceptron, and the entire network consists of multiple layers Perceptron, features are continuously extracted and compressed through multiple convolutional layers + activation functions (perceptrons), and finally relatively high-level features can be obtained, that is, the original features are concentrated step by step through the multi-layer neural network, and finally The obtained features are more reliable, and finally the last layer of features can be used to do various tasks: such as classification (such as judging the category), regression (such as returning to the detection frame), etc.
在本发明中，所述矾花识别AI模型的工作原理主要是：In the present invention, the operating principle of the alum flower recognition AI model is mainly:
1、通过多层卷积+激活层，提取图像的高级特征；1. Extract the advanced features of the image through multi-layer convolution + activation layer;
2、利用图像高级特征，通过回归任务，回归得出矾花的检测框(位置和大小)；2. Using the advanced features of the image, through the regression task, the detection frame (position and size) of the alum flower is obtained by regression;
3、利用图像高级特征，通过分类任务，分类得出矾花的类型(蓬松、片状、模糊矾花)。3. Using the advanced features of the image, through the classification task, the type of alum flower (fluffy, flaky, fuzzy alum flower) can be obtained by classification.
其训练过程如下：Its training process is as follows:
·收集大量矾花图像，并进进行人工打标(标注矾花位置、大小、类型)Collect a large number of alum flower images and carry out manual marking (marking the position, size and type of alum flowers)
·通过大量矾花数据，利用反向传播优化模型参数，训练矾花模型(多层神经网络)。·Through a large amount of alum flower data, use backpropagation to optimize the model parameters and train the alum flower model (multi-layer neural network).
图像采集识别单元11，用于利用水下摄像头采集矾花图像，利用训练好的所述矾花识别AI模型进行识别。The image collection and recognition unit 11 is used to collect images of alum flowers with an underwater camera, and use the trained AI model for recognition of alum flowers to perform recognition.
在本发明中，利用所述训练好的矾花识别AI模型能够检测拍摄图像中每个检出矾花类别、概率以及位置等属性。检出属性列表如下：In the present invention, the trained alum flower recognition AI model can detect attributes such as category, probability and position of each detected alum flower in the captured image. The list of checked out properties is as follows:
表1检测分割网络检出矾花相关属性表Table 1 The related attributes of alum flowers detected by the detection and segmentation network
Figure PCTCN2021119244-appb-000015
Figure PCTCN2021119244-appb-000015
Figure PCTCN2021119244-appb-000016
Figure PCTCN2021119244-appb-000016
量化评价指标检测单元12，用于对单帧图像进行区域划分为M×M的区域，分别对单一区域计算矾花量化指标，并对单一区域3类矾花量化指标进行组合得到单帧图像的量化评价指标，其中单一区域每一矾花类别的量化指标包括该类别在该区域的个数、置信平均、置信中位、面积平均以及面积中位。The quantitative evaluation index detection unit 12 is used to divide the single-frame image into M×M regions, calculate the alum flower quantification index for the single area respectively, and combine the three types of alum flower quantification indexes in the single area to obtain the single-frame image Quantitative evaluation indicators, wherein the quantitative indicators of each alum flower category in a single area include the number of the category in the area, confidence average, confidence median, area average, and area median.
由于拍摄图像区域内矾花分布不均匀且为了保留矾花分布信息，在量化矾花特征之前，对单帧图像进行区域划分为M×M的区域。以2×2为例，如图10所示，将整张图像以2×2的形式划分成如下4个区域，分别对单一区域计算矾花量化指标Due to the uneven distribution of alum flowers in the captured image area and in order to preserve the distribution information of alum flowers, before quantifying the characteristics of alum flowers, the single frame image was divided into M×M regions. Taking 2×2 as an example, as shown in Figure 10, the entire image is divided into the following 4 regions in the form of 2×2, and the quantitative indicators of alum flowers are calculated for a single region
单帧单一区域矾花量化指标计算：以图10左上区域为例来说明单帧单一区域矾花量化指标计算方式。如图11所示，对单一区域内，每一类矾花类别分别计算该类别在该区域的个数、置信平均、置信中位、面积平均以及面积中位。Calculation of quantitative indicators of alum flower in single frame and single area: take the upper left area of Figure 10 as an example to illustrate the calculation method of quantitative index of alum flower in single frame and single area. As shown in Figure 11, for each type of alum flower category in a single area, the number, confidence average, confidence median, area average, and area median of the category in the area are calculated respectively.
具体地，置信平均数、置信中位数计算方式：Specifically, the calculation method of confidence mean and confidence median:
Figure PCTCN2021119244-appb-000017
Figure PCTCN2021119244-appb-000017
其中，
Figure PCTCN2021119244-appb-000018
表示c类矾花的平均置信，
Figure PCTCN2021119244-appb-000019
表示c类矾花的置信中位数，N c表示该区域内c类矾花的个数。 in,
Figure PCTCN2021119244-appb-000018
Indicates the average confidence of class c alum flower,
Figure PCTCN2021119244-appb-000019
Indicates the confidence median of type c alum flowers, and N c indicates the number of type c alum flowers in this area.
面积平均数、面积中位数计算方式：Calculation method of area mean and area median:
Figure PCTCN2021119244-appb-000020
Figure PCTCN2021119244-appb-000020
其中，
Figure PCTCN2021119244-appb-000021
表示c类矾花的平均面积，
Figure PCTCN2021119244-appb-000022
表示c类矾花的面积中位数，A表示该 区域的面积。 in,
Figure PCTCN2021119244-appb-000021
Indicates the average area of c-type alum flowers,
Figure PCTCN2021119244-appb-000022
Indicates the median area of type c alum flowers, and A indicates the area of this area.
因此，单一区域每一矾花类别的量化指标：Therefore, the quantitative indicators for each type of alum in a single region:
Figure PCTCN2021119244-appb-000023
Figure PCTCN2021119244-appb-000023
其中，c代表矾花类别，i代表划分区域。Among them, c represents the category of alum flowers, and i represents the division area.
在本发明中，矾花各指标代表的物理意义：In the present invention, the physical significance that each index of alum flower represents:
(1)个数：当前区域某类检出矾花的个数；(1) Number: the number of a certain type of alum flowers detected in the current area;
(2)置信平均、置信中位：置信即矾花识别AI模型预测检出矾花的概率，表征检出矾花为当前类别的置信度；(2) Confidence average, confidence median: Confidence is the probability that the alum flower recognition AI model predicts the detection of alum flower, which represents the confidence that the detected alum flower is the current category;
(3)面积平均、面积中位：当前区域某类检出矾花面积之和相对于该区域的面积的比值，代表该区域该类矾花的密集程度。(3) Average area and median area: The ratio of the sum of the area of a certain type of alum flowers detected in the current area to the area of the area represents the density of this type of alum flowers in the area.
由上可知，单一区域每类矾花具有5个量化评价指标，根据矾花类别划分可得单一区域矾花指标共计5×3＝15个。It can be seen from the above that each type of alum flower in a single region has 5 quantitative evaluation indicators, and according to the classification of alum flower categories, a total of 5 × 3 = 15 alum flower indicators in a single region can be obtained.
单帧多区域矾花量化指标组合：将M×M的所有区域3类矾花量化指标进行组合可得到15M 2维的矾花量化指标，即为单帧图像的量化评价指标。 Combination of single-frame and multi-region alum quantitative indicators: Combining the three types of alum quantitative indicators in all regions of M×M can obtain 15M 2- dimensional alum quantitative indicators, which are quantitative evaluation indicators for single-frame images.
Figure PCTCN2021119244-appb-000024
Figure PCTCN2021119244-appb-000024
以图10的2×2划分为例，所提取的矾花量化指标个数为15×2×2＝60个。Taking the 2×2 division in Figure 10 as an example, the number of quantitative indicators of alum flowers extracted is 15×2×2=60.
矾花量化指标时间量化：由于原水数据1分钟更新一次，而由于摄像头FPS特征，1分钟约有360帧图像，若所有图像都参与矾花量化指标的计算，则会导致：(1)每帧图像都要经过矾花识别AI模型以及单帧多区域矾花量化评价的计算，计算复杂度过大；(2)相邻帧矾花量化指标相近，特征冗余。Time quantification of alum quantitative indicators: Since the raw water data is updated once a minute, and due to the FPS characteristics of the camera, there are about 360 frames of images in one minute. If all images participate in the calculation of alum quantitative indicators, it will result in: (1) every frame The images have to go through the calculation of the AI model of alum flower recognition and the quantitative evaluation of multi-region alum flower in a single frame, and the calculation complexity is too large; (2) the quantitative indicators of alum flower in adjacent frames are similar, and the features are redundant.
因此，在实际计算1分钟的矾花量化指标时，需要对时间量化，采用的方式：从1分钟内采集的图像等间隔选取F帧，计算单帧图像的矾花量化指标，最后F帧相同矾花量化指标求平均，即可得到时间量化后的矾花量化指标。Therefore, when actually calculating the quantitative index of alum flower for 1 minute, it is necessary to quantify the time. The method adopted is to select F frames at equal intervals from the images collected within 1 minute, and calculate the quantitative index of alum flower for a single frame image. Finally, the F frames are the same The quantitative index of the alum flower is averaged to obtain the quantitative index of the alum flower after time quantification.
实施例Example
实验数据：和水厂业务专家沟通，由于水厂方面净水***正在进行业务改造，混凝剂过量投放导致在沉淀池入水口处矾花，大多为颗粒状。因此，在初期相关参数确认的实验中，用4月24号～4月25号相对常规的数据。本实施例中以38个小时的数据作为训练集，以10个小时的数据作为测试集，在实验过程中，训练集：2280条，测试集：600条。Experimental data: Communicating with the business experts of the water plant, because the water purification system of the water plant is undergoing business transformation, the excessive use of coagulant has caused alum flowers at the water inlet of the sedimentation tank, most of which are granular. Therefore, relatively conventional data from April 24th to April 25th were used in the initial experiments to confirm relevant parameters. In this embodiment, 38 hours of data are used as a training set, and 10 hours of data are used as a test set. During the experiment, the training set: 2280 records, and the test set: 600 records.
单帧图像区域划分M×M实验：在实验过程中，将单帧图像分别划分为1×1、2×2、3×3、4×4、5×5、6×6、7×7、10×10的区域，建立支持向量回归模型(只使用矾花量化指标建模)，在相同测试集，各模型的得分如下表2：Single-frame image area division M×M experiment: During the experiment, the single-frame image was divided into 1×1, 2×2, 3×3, 4×4, 5×5, 6×6, 7×7, In the 10×10 area, a support vector regression model was established (using only the quantitative indicators of alum flowers to model). In the same test set, the scores of each model are as follows in Table 2:
表2单帧图像区域划分M×M实验Table 2 M×M experiment of single frame image area division
SVR SVR 1×11×1 2×22×2 3×33×3 5×55×5 6×66×6 7×77×7 10×1010×10
测试集test set 600600 600600 600600 600600 600600 600600 600600
模型得分model score 0.55730.5573 0.59280.5928 0.55240.5524 0.54440.5444 0.55760.5576 0.52740.5274 0.46480.4648
矾花量化指标时间量化参数F帧实验：在实验过程中，每分钟等间隔分别采样10、25、30、40、50、60、70帧图像进行矾花量化指标计算，建立支持向量回归模型(只使用矾花量化指标建模)，在相同测试集，各模型得分如下表3：Quantitative index time quantification parameter F frame experiment of alum flower: During the experiment, 10, 25, 30, 40, 50, 60, and 70 frames of images were sampled at equal intervals every minute to calculate the quantitative index of alum flower, and a support vector regression model was established ( Only use the quantitative indicators of alum flowers to model), in the same test set, the scores of each model are as follows in Table 3:
表3矾花量化指标时间量化参数F帧实验Table 3 Alum flower quantification index time quantification parameter F frame experiment
SVR SVR 1010 2525 3030 4040 5050 6060 7070
测试集test set 600600 600600 600600 600600 600600 600600 600600
模型得分model score 0.48510.4851 0.59280.5928 0.60560.6056 0.60310.6031 0.60590.6059 0.61380.6138 0.59750.5975
分析：(1)对于单帧图像区域划分实验来说，虽然划分区域为2×2和5×5的得分都在0.4左右，但是相对于计算量和精度的权衡，在本项目中，将单帧图像划分为2×2进行矾花量化指标计算；(2)对于矾花量化指标时间量化参数实验，考虑到算法实时性，选择每分钟采集25帧图像，计算时间量化后的矾花量化指标。Analysis: (1) For the single-frame image area division experiment, although the scores of dividing the area into 2×2 and 5×5 are both around 0.4, compared to the trade-off between calculation amount and accuracy, in this project, the single The frame image is divided into 2×2 to calculate the quantitative index of alum flower; (2) For the time quantization parameter experiment of the quantitative index of alum flower, considering the real-time performance of the algorithm, 25 frames of images are collected per minute, and the quantitative index of alum flower after calculating the time quantization .
由以上分析，建立多模态数据源滞后回归沉后水浊度模型除了预测大约2小时之后的沉后水浊度之外，还可以根据训练好的模型评估矾花量化指标是否合理。根据本发明所建立模型可计算出矾花量化各个指标对于沉后水浊度的归一化重要程度，并给出量化指标，如图13所示。Based on the above analysis, in addition to predicting the turbidity of submerged water about 2 hours later, the establishment of a multimodal data source lag regression model of turbidity of submerged water can also evaluate whether the quantitative indicators of alum flowers are reasonable based on the trained model. According to the model established in the present invention, the normalized importance of various indicators of alum flower quantification to the turbidity of the sinking water can be calculated, and the quantitative indicators are given, as shown in FIG. 13 .
从图13可以看出，在各个划分区域中蓬松矾花的量化指标的重要程度较高(例如：区域2_蓬松矾花_平均置信度：0.17，区域2_蓬松矾花_个数：0.13等)，而片状矾花和模糊蓬松对于沉后水浊度预测重要性较低，和矾花业务分析的定性结论相吻合。进一步地，对所有区域不同类别矾花重要性进行分析，如图14所示。结合图13与图14，从定量的角度分析，蓬松矾花的量化指标对于沉后水浊度的预测比较重要，和业务分析的定性结论一致，验证了矾花细粒分类和量化指标的合理性。As can be seen from Figure 13, the importance of the quantitative index of fluffy alum flowers in each division area is higher (for example: area 2_fluffy alum flowers_average confidence: 0.17, area 2_fluffy alum flowers_number: 0.13 etc.), while the flaky alum flowers and fuzzy fluffy are less important for the prediction of the turbidity of the sinking water, which is consistent with the qualitative conclusion of the alum flower business analysis. Further, the importance of different types of alum flowers in all regions was analyzed, as shown in Figure 14. Combining Figure 13 and Figure 14, from a quantitative point of view, the quantitative index of the fluffy alum flower is more important for the prediction of the turbidity of the sinking water, which is consistent with the qualitative conclusion of the business analysis, and verifies the reasonableness of the classification of the fine-grained alum flower and the quantitative index sex.
综上所述，对所有区域每分钟各类别矾花的量化指标(矾花个数、置信平 均、置信中位、面积平均以及面积中位)求取平均，归一化之后显示在Web界面的柱状图上，通过各区域矾花量化指标的定量重要性指标可知，当蓬松矾花个数越多，面积越大时，矾花的混凝效果越好。To sum up, the quantitative indicators (number of alum flowers, confidence average, confidence median, area average, and area median) of each category of alum flowers per minute in all areas are averaged, and displayed on the web interface after normalization. On the histogram, it can be seen from the quantitative importance indicators of the quantitative indicators of alum flowers in each region that the coagulation effect of alum flowers is better when the number of fluffy alum flowers is larger and the area is larger.
上述实施例仅例示性说明本发明的原理及其功效，而非用于限制本发明。任何本领域技术人员均可在不违背本发明的精神及范畴下，对上述实施例进行修饰与改变。因此，本发明的权利保护范围，应如权利要求书所列。The above-mentioned embodiments only illustrate the principles and effects of the present invention, but are not intended to limit the present invention. Any person skilled in the art can modify and change the above-mentioned embodiments without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention should be listed in the claims.
工业实用性Industrial Applicability
所属领域技术人员根据上文的记载容易得知，本发明技术方案适合在工业中制造并在生产、生活中使用，因此本发明具备工业实用性。Those skilled in the art can easily know from the above description that the technical solution of the present invention is suitable for industrial manufacture and use in production and daily life, so the present invention has industrial applicability.

Claims (11)

  1. 一种利用图像识别技术分析矾花特征的方法，包括如下步骤：A method utilizing image recognition technology to analyze the characteristics of alum flowers, comprising the steps of:
    步骤S1，利用水下摄像头采集矾花图像，对采集图像打标，生成训练数据，建立矾花识别AI模型，利用打标数据作为监督信号，对矾花识别AI模型进行自主训练，获得训练好的矾花识别AI模型；Step S1, use the underwater camera to collect alum flower images, mark the collected images, generate training data, establish an AI model for alum flower recognition, use the marking data as a supervisory signal, conduct independent training on the alum flower recognition AI model, and obtain a well-trained Alum flower recognition AI model;
    步骤S2，利用水下摄像头采集矾花图像，利用训练好的所述矾花识别AI模型进行识别；Step S2, using the underwater camera to collect images of alum flowers, and using the trained AI model for identifying alum flowers to identify them;
    步骤S3，对单帧图像进行区域划分为M×M的区域，分别对单一区域计算矾花量化指标，并对单一区域各类矾花量化指标进行组合得到单帧图像的量化评价指标。In step S3, the single frame image is divided into M×M regions, and the quantitative index of alum flower is calculated for each single area, and the quantitative evaluation index of the single frame image is obtained by combining various quantitative indexes of alum flower in the single area.
  2. 如权利要求1所述的一种利用图像识别技术分析矾花特征的方法，其特征在于，步骤S1进一步包括：A kind of method utilizing image recognition technology to analyze alum flower feature as claimed in claim 1, is characterized in that, step S1 further comprises:
    步骤S100，利用水下摄像头采集其景深范围内的矾花图像；Step S100, using the underwater camera to collect images of alum flowers within its depth of field;
    步骤S101，确定各类别矾花标注标准，对采集图像打标，生成训练数据；Step S101, determining the labeling standards for each category of alum flowers, marking the collected images, and generating training data;
    步骤S102，建立多视角矾花检测分割网络的矾花识别AI模型，利用打标数据作为监督信号，对多视角矾花检测分割网络进行训练，拟合标注数据。In step S102, an AI model for alum flower recognition of a multi-view alum flower detection and segmentation network is established, and the marking data is used as a supervisory signal to train the multi-view alum flower detection and segmentation network to fit the labeled data.
  3. 如权利要求2所述的一种利用图像识别技术分析矾花特征的方法，其特征在于：于步骤S101中，依据矾花的相关业务以及拍摄图像，确定三种类别矾花的标注标准，根据确定的标注标准对采集图像打标，对于在图像上无法确定蓬松矾花和片状矾花、矾花和模糊蓬松的图像，添加备注数据。A method for analyzing the characteristics of alum flowers using image recognition technology as claimed in claim 2, characterized in that: in step S101, according to the related business of alum flowers and the captured images, the labeling standards of three types of alum flowers are determined, according to The determined labeling standard is used to mark the collected images, and for images where fluffy alum flowers and flaky alum flowers, alum flowers and fuzzy fluffy images cannot be determined on the images, add remark data.
  4. 如权利要求3所述的一种利用图像识别技术分析矾花特征的方法，其特征在于：于步骤S102中，所述矾花识别AI模型采用多层神经网络，通过多层卷积与激活层，提取图像的高级特征，并利用所提取的图像高级特征，通过回归任务回归得出矾花的检测框，以及利用图像高级特征，通过分类任务分类得出矾花的类型。A method for analyzing the characteristics of alum flowers using image recognition technology as claimed in claim 3, characterized in that: in step S102, the alum flower recognition AI model adopts a multi-layer neural network, through multi-layer convolution and activation layers , extract the advanced features of the image, and use the extracted advanced features of the image to obtain the detection frame of the alum flower through regression task regression, and use the advanced image feature to classify the type of alum flower through the classification task.
  5. 如权利要求4所述的一种利用图像识别技术分析矾花特征的方法，其 特征在于：于步骤S102中，通过大量矾花数据，利用反向传播优化模型参数，训练所述矾花识别AI模型。A method for analyzing the characteristics of alum flowers using image recognition technology as claimed in claim 4, characterized in that: in step S102, through a large amount of alum flower data, use back propagation to optimize model parameters, and train the alum flower recognition AI Model.
  6. 如权利要求3所述的一种利用图像识别技术分析矾花特征的方法，其特征在于：于步骤S2中，利用所述训练好的矾花识别AI模型识别采集图像，检出每个矾花类别、概率以及位置属性。A method for analyzing the characteristics of alum flowers using image recognition technology as claimed in claim 3, characterized in that: in step S2, using the trained alum flower recognition AI model to identify and collect images, and detect each alum flower Category, probability, and location attributes.
  7. 如权利要求5所述的一种利用图像识别技术分析矾花特征的方法，其特征在于：所述单一区域每一矾花类别的量化指标包括该类别在该区域的个数、置信平均、置信中位、面积平均以及面积中位。A method for analyzing the characteristics of alum flowers using image recognition technology as claimed in claim 5, characterized in that: the quantitative index of each alum flower category in the single area includes the number of the category in the area, the confidence average, the confidence Median, area mean, and area median.
  8. 如权利要求6所述的一种利用图像识别技术分析矾花特征的方法，其特征在于：于步骤S3中，将M×M的所有区域三类矾花量化指标进行组合得到15M 2维的矾花量化指标，即为单帧图像的量化评价指标。 A method for analyzing the characteristics of alum flowers using image recognition technology as claimed in claim 6, characterized in that: in step S3, the quantitative indicators of three types of alum flowers in all regions of M×M are combined to obtain 15M 2- dimensional alum flowers Flower quantitative index is the quantitative evaluation index of a single frame image.
  9. 如权利要求7所述的一种利用图像识别技术分析矾花特征的方法，其特征在于：于步骤S3中，从1分钟内采集的图像等间隔选取F帧，计算单帧图像的矾花量化指标，然后将F帧相同矾花量化指标求平均，得到时间量化后的矾花量化指标。A method for analyzing the characteristics of alum flowers using image recognition technology as claimed in claim 7, characterized in that: in step S3, F frames are selected at equal intervals from the images collected within 1 minute, and the quantification of alum flowers in a single frame image is calculated index, and then calculate the average of the same flower quantization indicators in F frame to obtain the flower quantization index after time quantization.
  10. 一种利用图像识别技术分析矾花特征的装置，包括：A device for analyzing the characteristics of alum flowers using image recognition technology, comprising:
    矾花识别AI模型构建及训练单元，用于利用水下摄像头采集矾花图像，对采集图像打标，生成训练数据，建立矾花识别AI模型，利用打标数据作为监督信号，对矾花识别AI模型进行自主训练，获得训练好的矾花识别AI模型；Alum flower recognition AI model construction and training unit, used to use underwater camera to collect alum flower images, mark the collected images, generate training data, establish alum flower recognition AI model, use the marking data as a supervisory signal, and identify alum flower The AI model is trained independently to obtain the trained AI model for alum flower recognition;
    图像采集识别单元，用于利用水下摄像头采集矾花图像，利用训练好的所述矾花识别AI模型进行识别；The image collection and recognition unit is used to collect the alum flower image by using the underwater camera, and use the trained AI model to identify the alum flower;
    量化评价指标检测单元，用于对单帧图像进行区域划分为M×M的区域，分别对单一区域计算矾花量化指标，并对单一区域三类矾花量化指标进行组合得到单帧图像的量化评价指标。Quantitative evaluation index detection unit, which is used to divide a single frame image into M×M areas, calculate the quantitative index of alum flower for a single area, and combine the three types of alum flower quantitative indicators in a single area to obtain the quantification of a single frame image evaluation index.
  11. 如权利要求9所述的一种利用图像识别技术分析矾花特征的装置，其特征在于，所述量化评价指标检测单元从1分钟内采集的图像等间隔选取F帧， 计算单帧图像的矾花量化指标，然后将F帧相同矾花量化指标求平均，得到时间量化后的矾花量化指标。A kind of device that utilizes image recognition technology to analyze the feature of alum flower as claimed in claim 9, it is characterized in that, described quantitative evaluation index detection unit selects F frames at equal intervals from images collected within 1 minute, calculates the alum of single frame image The flower quantization index is calculated, and then the same flower quantization index of the F frame is averaged to obtain the flower quantization index after time quantization.
PCT/CN2021/119244 2021-08-30 2021-09-18 Method and apparatus for analyzing alum floc feature by using image recognition technology WO2023029117A1 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111001467.4A CN113688751A (en) 2021-08-30 2021-08-30 Method and device for analyzing alum blossom characteristics by using image recognition technology
CN202111001467.4 2021-08-30

Publications (1)

Publication Number Publication Date
WO2023029117A1 true WO2023029117A1 (en) 2023-03-09

Family

ID=78583803

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/119244 WO2023029117A1 (en) 2021-08-30 2021-09-18 Method and apparatus for analyzing alum floc feature by using image recognition technology

Country Status (2)

Country Link
CN (1) CN113688751A (en)
WO (1) WO2023029117A1 (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116229518A (en) * 2023-03-17 2023-06-06 百鸟数据科技(北京)有限责任公司 Bird species observation method and system based on machine learning
CN117929375A (en) * 2024-03-21 2024-04-26 武汉奥恒胜科技有限公司 Water quality detection method and water quality detector based on image processing

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115375009B (en) * 2022-07-28 2023-09-22 上海城市水资源开发利用国家工程中心有限公司 Method for establishing intelligent monitoring linkage system for coagulation
CN117315454B (en) * 2023-11-29 2024-03-12 河北中瀚水务有限公司 Evaluation method, device and system for flocculation reaction process

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111833369A (en) * 2020-07-21 2020-10-27 中冶赛迪重庆信息技术有限公司 Alum image processing method, system, medium and electronic device
CN112101352A (en) * 2020-09-10 2020-12-18 广州深视未来智能科技有限责任公司 Underwater alumen ustum state identification method and monitoring device, computer equipment and storage medium
CN112441654A (en) * 2020-11-02 2021-03-05 广州晋合水处理设备有限公司 Control system and method suitable for coagulating sedimentation

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8811708B2 (en) * 2009-04-15 2014-08-19 Koninklijke Philips N.V. Quantification of medical image data
JP6789515B2 (en) * 2017-03-24 2020-11-25 株式会社サイバーブレーン Programs, devices, systems and methods for generating image information to be photographed
US11232330B2 (en) * 2018-02-13 2022-01-25 Slingshot Aerospace, Inc. Adaptive neural network selection to extract particular results
CN111242918B (en) * 2020-01-10 2023-01-24 深圳信息职业技术学院 Image segmentation method and system based on Kalman filtering and Markov random field

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111833369A (en) * 2020-07-21 2020-10-27 中冶赛迪重庆信息技术有限公司 Alum image processing method, system, medium and electronic device
CN112101352A (en) * 2020-09-10 2020-12-18 广州深视未来智能科技有限责任公司 Underwater alumen ustum state identification method and monitoring device, computer equipment and storage medium
CN112441654A (en) * 2020-11-02 2021-03-05 广州晋合水处理设备有限公司 Control system and method suitable for coagulating sedimentation

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116229518A (en) * 2023-03-17 2023-06-06 百鸟数据科技(北京)有限责任公司 Bird species observation method and system based on machine learning
CN116229518B (en) * 2023-03-17 2024-01-16 百鸟数据科技(北京)有限责任公司 Bird species observation method and system based on machine learning
CN117929375A (en) * 2024-03-21 2024-04-26 武汉奥恒胜科技有限公司 Water quality detection method and water quality detector based on image processing
CN117929375B (en) * 2024-03-21 2024-06-04 武汉奥恒胜科技有限公司 Water quality detection method and water quality detector based on image processing

Also Published As

Publication number Publication date
CN113688751A (en) 2021-11-23

Similar Documents

Publication Publication Date Title
WO2023029117A1 (en) Method and apparatus for analyzing alum floc feature by using image recognition technology
CN109300111B (en) Chromosome recognition method based on deep learning
CN107742099A (en) A kind of crowd density estimation based on full convolutional network, the method for demographics
CN111966053A (en) Intelligent flocculant decision making system
CN108363953B (en) Pedestrian detection method and binocular monitoring equipment
CN108647643A (en) A kind of packed tower liquid flooding state on-line identification method based on deep learning
Tian et al. CASDD: Automatic surface defect detection using a complementary adversarial network
CN108931621B (en) Zinc ore grade soft measurement method based on process texture characteristics
CN111274886A (en) Deep learning-based pedestrian red light violation analysis method and system
CN112184709A (en) Water body turbidity prediction method and system based on artificial intelligence
CN114170686A (en) Elbow bending behavior detection method based on human body key points
CN113077423B (en) Laser selective melting pool image analysis system based on convolutional neural network
CN107633527A (en) Target tracking method and device based on full convolutional neural networks
Shishkin et al. Implementation of yolov5 for detection and classification of microplastics and microorganisms in marine environment
Tian et al. Feature fusion–based preprocessing for steel plate surface defect recognition
CN113379247B (en) Modeling method and system for enterprise potential safety hazard tracking model
CN111258996B (en) Product quality multi-source depth fusion forecasting method for industrial big data
Kumari et al. Driver distraction analysis using convolutional neural networks
TW202326114A (en) Fault or defect detection system for long-code zipper capable of meeting the requirement of continuous detection of the production assembly line having transmission speed at 300 m/min or lower, and has high detection precision
CN111767815A (en) Tunnel water leakage identification method
Venkatesh et al. Deep learning for macro-nutrient deficiency identification in the groundnut plants
CN110728310A (en) Target detection model fusion method and system based on hyper-parameter optimization
CN110189360A (en) A kind of recognition and tracking method of pair of specific objective
Guo et al. Research on position recognition system of gantry hoisting based on machine vision
CN116597355B (en) Small micro water body state real-time monitoring method, system and device based on vision system

Legal Events

Date Code Title Description
NENP Non-entry into the national phase

Ref country code: DE