CN112950586A - LF furnace steel slag infrared identification method and system - Google Patents

LF furnace steel slag infrared identification method and system Download PDF

Info

Publication number
CN112950586A
CN112950586A CN202110228878.0A CN202110228878A CN112950586A CN 112950586 A CN112950586 A CN 112950586A CN 202110228878 A CN202110228878 A CN 202110228878A CN 112950586 A CN112950586 A CN 112950586A
Authority
CN
China
Prior art keywords
steel slag
infrared
furnace
image data
image
Prior art date
Legal status (The legal status 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 status listed.)
Pending
Application number
CN202110228878.0A
Other languages
Chinese (zh)
Inventor
刘凯
李小平
曹树森
彭宏亮
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Pangang Group Panzhihua Iron and Steel Research Institute Co Ltd
Original Assignee
Pangang Group Panzhihua Iron and Steel Research Institute Co Ltd
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 Pangang Group Panzhihua Iron and Steel Research Institute Co Ltd filed Critical Pangang Group Panzhihua Iron and Steel Research Institute Co Ltd
Priority to CN202110228878.0A priority Critical patent/CN112950586A/en
Publication of CN112950586A publication Critical patent/CN112950586A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0004—Industrial image inspection
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/90—Dynamic range modification of images or parts thereof
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10048—Infrared image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]

Landscapes

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

Abstract

The application discloses an LF furnace steel slag infrared identification method, which comprises the steps of carrying out infrared detection on steel slag on the surface of a steel ladle of an LF furnace and bare molten steel in the steel ladle to obtain image data; preprocessing the image data; performing data augmentation on the preprocessed image data to obtain a training data set and a test data set; training a SegNet network model by using the training data set to obtain a deep learning model; and testing the test data set by using the deep learning model, and identifying the steel slag in the image data. The method can quickly and accurately adjust the flow of argon blown from the bottom of the molten steel, has strong anti-interference capability, ensures the impurities in the molten steel to fully float, and stably controls the cleanliness of the molten steel. The application also discloses an LF furnace steel slag infrared identification system which has the same advantages as the method.

Description

LF furnace steel slag infrared identification method and system
Technical Field
The invention belongs to the technical field of steel making, and particularly relates to an LF furnace steel slag infrared identification method and system.
Background
In modern steelmaking production, bottom blowing argon stirring treatment is needed on molten steel, the reason for bottom blowing is that after an LF furnace is subjected to operations such as deoxidation and desulfurization, a large amount of calcium aluminate inclusions can be suspended in the molten steel, once the inclusions enter a casting blank, serious adverse effects can be caused on the surface and internal quality of a subsequent rolled material, and bottom blowing argon treatment on the molten steel can promote the inclusions in the molten steel to float upwards, so that the components and temperature of the molten steel are more uniform, the molten steel is purer, the cleanliness of the molten steel is improved, and a billet with higher quality is obtained.
And when argon is blown at the bottom, the flow of argon is mainly adjusted according to the condition of steel slag, in the prior art, the condition of the steel slag is observed only by a manual mode, the randomness of the mode is high, the quality of molten steel is unstable, the quality of steel billets is influenced, the service life of steel ladles is influenced, and argon is blown at the bottom is wasted. Therefore, the existing manual visual inspection method lacks scientific basis, and is greatly influenced by subjective factors, while the commonly adopted common image processing method needs to carry out preprocessing such as denoising and the like on an infrared image, so that image distortion is easily caused, the detection is carried out by setting a color threshold range, the accuracy of a detection result is not high, a recognition error occurs when the color is slightly changed, and the defects reduce the accuracy of the detection, so that the detection method cannot be used for on-site steel slag detection.
Disclosure of Invention
In order to solve the problems, the invention provides an LF furnace steel slag infrared identification method and system, which can quickly and accurately adjust the flow of argon blown at the bottom of molten steel, have strong anti-interference capability and ensure that impurities in the molten steel fully float upwards, thereby stably controlling the cleanliness of the molten steel.
The invention provides an LF furnace steel slag infrared identification method which comprises the following steps:
carrying out infrared detection on steel slag on the surface of a steel ladle of the LF furnace and bare molten steel in the steel ladle to obtain image data;
preprocessing the image data;
performing data augmentation on the preprocessed image data to obtain a training data set and a test data set;
training a SegNet network model by using the training data set to obtain a deep learning model;
and testing the test data set by using the deep learning model, and identifying the steel slag in the image data.
Preferably, in the infrared identification method for the steel slag of the LF furnace, an infrared thermal imager arranged above the LF furnace is used for carrying out infrared detection on the steel slag on the surface of the steel ladle of the LF furnace and the bare molten steel in the steel ladle.
Preferably, in the LF furnace steel slag infrared identification method, the preprocessing the image data includes:
and performing enhancement and noise reduction processing on the image data.
Preferably, in the LF furnace steel slag infrared identification method, the data augmentation of the preprocessed image data includes:
and carrying out data augmentation on the preprocessed image data under a small sample amount through a generative countermeasure network.
Preferably, in the LF furnace steel slag infrared recognition method, before training the SegNet network model using the training data set, the method further includes building the SegNet network model:
establishing an input layer, wherein the input layer inputs the slices of the infrared steel slag detection picture;
building convolution layers, performing feature extraction and feature mapping by using convolution kernels, and extracting image features by using 4 groups of convolution layers;
and establishing a deconvolution layer, and performing up-sampling operation on the feature mapping image by using deconvolution to restore the image to the size of the original image.
The invention provides an LF furnace steel slag infrared identification system, which comprises:
the image data acquisition device is used for carrying out infrared detection on steel slag on the surface of the steel ladle of the LF furnace and bare molten steel in the steel ladle to obtain image data;
preprocessing means for preprocessing the image data;
the data amplification device is used for performing data amplification on the preprocessed image data to obtain a training data set and a test data set;
the training device is used for training the SegNet network model by utilizing the training data set to obtain a deep learning model;
and the steel slag identification device is used for testing the test data set by using the deep learning model and identifying the steel slag in the image data.
Preferably, in the LF furnace steel slag infrared identification system, the image data acquisition device is a thermal infrared imager installed above the LF furnace.
Preferably, in the LF furnace steel slag infrared identification system, the preprocessing device is specifically configured to perform enhancement and noise reduction processing on the image data.
Preferably, in the LF furnace steel slag infrared identification system, the data amplification device is specifically configured to perform data amplification on the preprocessed image data through a generative countermeasure network in a small sample amount.
Preferably, in the LF furnace steel slag infrared recognition system, the system further comprises a SegNet network model building device, specifically comprising an input layer building unit for building an input layer, wherein the input layer inputs the slices of the infrared steel slag detection picture;
the convolutional layer establishing unit is used for establishing a convolutional layer, performing feature extraction and feature mapping by using the convolutional layer, and extracting image features by using 4 groups of convolutional layers;
and the deconvolution layer establishing unit is used for establishing a deconvolution layer and performing up-sampling operation on the feature mapping image by using deconvolution so as to restore the image to the size of the original image.
According to the description, the LF furnace steel slag infrared identification method provided by the invention comprises the following steps of carrying out infrared detection on steel slag on the surface of a steel ladle of the LF furnace and bare molten steel in the steel ladle to obtain image data; preprocessing the image data; performing data augmentation on the preprocessed image data to obtain a training data set and a test data set; training a SegNet network model by using the training data set to obtain a deep learning model; the deep learning model is used for testing the test data set, and steel slag in the image data is identified, so that the flow of argon blown from the bottom of molten steel can be quickly and accurately adjusted, the anti-interference capability is high, the impurities in the molten steel can be ensured to float sufficiently, and the cleanliness of the molten steel can be stably controlled. The LF furnace steel slag infrared identification system provided by the invention has the same advantages as the LF furnace steel slag infrared identification method.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the provided drawings without creative efforts.
FIG. 1 is a schematic diagram of an embodiment of an LF furnace steel slag infrared identification method provided by the invention;
fig. 2 is a schematic diagram of a generative countermeasure structure in another embodiment of the LF furnace steel slag infrared identification method provided in the present application;
FIG. 3 is a schematic diagram of the overall architecture of an improved SegNet model;
FIG. 4 is a schematic diagram of a test result of a molten steel infrared image of a test set;
fig. 5 is a schematic diagram of an embodiment of an LF furnace steel slag infrared identification system provided by the present invention.
Detailed Description
The core of the invention is to provide the LF furnace steel slag infrared identification method and the LF furnace steel slag infrared identification system, which can quickly and accurately adjust the flow of argon blown at the bottom of molten steel, have strong anti-interference capability, ensure the full floating of impurities in the molten steel and stably control the cleanliness of the molten steel.
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Fig. 1 shows an embodiment of an LF furnace steel slag infrared identification method provided by the present invention, where fig. 1 is a schematic diagram of an embodiment of an LF furnace steel slag infrared identification method provided by the present invention, and the method may include the following steps:
s1: carrying out infrared detection on steel slag on the surface of a steel ladle of the LF furnace and bare molten steel in the steel ladle to obtain image data;
it should be noted that infrared thermal imaging is used as a non-contact detection means, has the advantages of rapidness, no damage, non-contact, no need of coupling, real-time performance, large detection range, intuition and the like, and can clearly observe an object to be monitored at night without light or in a severe environment of steel mill waste gas and smoke dust. The infrared thermal imaging technology can improve the detection performance, can realize the quick, real-time and accurate detection of the steel ladle, prevent the interference of on-site light and dust, achieve the effect of stably obtaining the original image in the steel ladle, and transmit the obtained infrared image data to a subsequent preprocessing device.
S2: preprocessing image data;
it should be noted that the preprocessing process may include, but is not limited to, image enhancement and noise reduction processing. Specifically, the method includes the steps of firstly calculating a multi-contrast combined image of absolute contrast, change contrast, normalized contrast, standard contrast and difference absolute contrast of an acquired infrared image to enhance a molten steel image in the infrared image, then denoising by using a wavelet transform threshold to remove uneven heating factors, noise and deformity in the image, wherein the infrared imaging system is easily interfered and uneven heating exists in the image Improper reconstruction methods or noise and malformations introduced by the external environment.
S3: performing data augmentation on the preprocessed image data to obtain a training data set and a test data set;
it should be noted that this data augmentation process is mainly used to add a training data set, diversify the data set as much as possible, and make the trained model have a stronger generalization capability, including horizontal and vertical flipping, rotation, scaling, clipping, shearing, translation, contrast, color dithering, noise, etc., and a large amount of image data is obtained after data augmentation, and then these data are divided into two parts, one part is used to train the SegNet network model, and the other part is used to test whether the model result has reached the expected standard after the model has been trained, i.e., whether the result is accurate.
S4: training the SegNet network model by utilizing a training data set to obtain a deep learning model;
it should be noted that the SegNet Network model mainly consists of an encoding Network (Encoder Network), a decoding Network (Decoder Network) and a Pixel-by-Pixel Classification Layer, and each convolution Layer is followed by a Batch Normalization Layer and a ReLU activation function, wherein the Batch Normalization is performed on data, that is, the data satisfies a normal distribution with a mean value of 0 and a variance of 1, and the main function of the SegNet Network model is to alleviate a gradient disappearance/explosion phenomenon in DNN training and accelerate the training speed of the model, the encoding Network converts a high-dimensional vector into a low-dimensional vector to realize low-dimensional extraction of high-dimensional features, and although the encoding Network can capture more translation invariance features through multiple maximum pooling operations, the encoding Network can also lose the important segmentation of more boundary information of characteristic maps and the like, so that the maximum pooling index information is recorded simultaneously in the pooling process, the position of the maximum characteristic value is stored, then the input characteristic graph is sampled by utilizing the maximum pooling index information, so that the boundary information can be stored, the decoding network utilizes the maximum pooling index information of the corresponding characteristic layer stored when the encoder performs down-sampling to map the characteristic diagram with low resolution to the characteristic diagram with high spatial resolution, the reconstruction from the low-dimensional vector to the high-dimensional vector is realized, the repeated use of the maximal pooling index in the decoding process has the advantages of optimizing the boundary profile description, reducing the number of parameters and training end to end, the upsampling mode can be applied to any encoding-decoding network, the high-dimensional feature representation vector is output at the last layer of decoder, and is used as the input of the trainable Softmax classifier, the SegNet network model is a full convolution network capable of performing pixel-level image segmentation, and can reconstruct the infrared steel slag detection problem into an end-to-end multi-classification problem.
S5: and testing the test data set by using the deep learning model, and identifying the steel slag in the image data.
It should be noted that the strong learning ability of the SegNet deep learning network is used for classifying the steel slag pixel points and the non-steel slag pixel points pixel by pixel, so that the limitation that the traditional steel slag image detection manually extracts features and an algorithm is designed according to specific problems by means of manual experience is broken. The method is simple to operate, strong in robustness and high in accuracy, meets the real-time requirement of industrial steel slag image detection, and through experiments, the infrared steel slag detection accuracy can reach 98.5% by adopting a SegNet-based deep learning method.
According to the description, in the embodiment of the LF furnace steel slag infrared identification method provided by the invention, the steel slag on the surface of the steel ladle of the LF furnace and the bare molten steel in the steel ladle are subjected to infrared detection to obtain image data; preprocessing image data; performing data augmentation on the preprocessed image data to obtain a training data set and a test data set; training the SegNet network model by utilizing a training data set to obtain a deep learning model; the deep learning model is used for testing the test data set, steel slag in the image data is identified, the percentage of the steel slag is identified more accurately, and therefore the more accurate argon demand can be corresponded, the flow of argon blown at the bottom of molten steel can be adjusted rapidly and accurately, the anti-interference capability is strong, the sufficient floating of impurities in the molten steel is guaranteed, and the cleanliness of the molten steel is controlled stably.
In a specific embodiment of the LF furnace steel slag infrared identification method, an infrared thermal imager installed above the LF furnace may be used to perform infrared detection on steel slag on the surface of a steel ladle of the LF furnace and bare molten steel in the steel ladle. It should be noted that the infrared thermal imager uses an infrared detector and an optical imaging objective to receive an infrared radiation energy distribution image of a measured object and reflect the infrared radiation energy distribution image onto a photosensitive element of the infrared detector, so as to obtain an infrared thermal image, and the thermal image corresponds to a thermal distribution field on the surface of an object. Infrared thermal imagers convert the invisible infrared energy emitted by an object into visible thermal images, the different colors on the thermal images representing the different temperatures of the object being measured. The non-contact infrared detection technology can be used for rapidly, accurately, conveniently and visually displaying the distribution of the surface temperature field of the object to be measured, measuring the surface temperature of the object, rapidly measuring the surface temperature reading of the object without directly contacting the surface of the object to be measured, reliably measuring the surface temperature of the object which is hot, dangerous or difficult to contact, and enabling the thermal infrared imager to be very fast in measurement speed and visually and continuously measure the temperature change of the surface of the object. Of course, this is only one preferred embodiment, and other types of infrared detection devices may be used according to actual needs, and are not limited herein.
In another specific embodiment of the LF furnace steel slag infrared identification method, the data augmentation of the preprocessed image data may include:
and carrying out data augmentation on the preprocessed image data under a small sample amount through a generative countermeasure network. Specifically, in the machine learning and training process, the problem that the generalization performance (general-purpose capability and anti-interference capability) of the trained model is poor due to insufficient data volume (i.e., the samples are not comprehensive enough) is often encountered, at this time, the technology of artificially expanding the data to generate more kinds of data (more in line with the actual situation) is data expansion, and the acquired LF furnace molten steel infrared image data is subjected to data expansion through a Generative Adaptive Network (GAN) to generate more infrared images, so that the infrared image data set of the LF furnace molten steel is expanded, and thus a large number of data samples required by convolutional neural network training are obtained, and the problem that the convolutional neural network is difficult to train by a small sample data set is solved. Referring to fig. 2, fig. 2 is a schematic diagram of a generative countermeasure structure in another embodiment of the LF furnace steel slag infrared identification method provided in the present application, where there are two networks, namely, a generative network g (generator) and a discriminant network d (discriminator). The generation network receives a random noise input (conforming to simple distribution such as Gaussian distribution or even distribution), and outputs a picture through the noise, wherein the picture is marked as G (z), the input of the network is judged to be x, x represents a picture, and D (x) represents the probability that x is a real picture, and the functions of the two models are respectively: g is responsible for generating pictures, receiving a random noise z, generating the pictures through the noise, marking the generated pictures as G (z), D is responsible for judging whether a picture is real or not, the input of the D is x, x represents a picture, D (x) represents the probability that x is the real picture, if the x is 1, the probability that x is the real picture is 100 percent, and the output of D is 0, the picture cannot be real (the real example is from the data set, and the forged example is from the generation model). In the training process, the object of generating the model G is to generate pictures which are similar to original data as much as possible to cheat the discriminant model D, the object of the discriminant model D is to distinguish the pictures generated by the generated model G from the real pictures as much as possible, so that a generator tries to cheat a discriminator, the discriminator tries not to cheat by the generator, the two models are alternately optimized and trained and mutually promoted, G and D form a dynamic game, and in the dynamic game process, the whole network is continuously optimized; when the result of the discrimination by the discriminator for an input is 0.5, G can generate enough pictures G (z) to be "spurious", and it is difficult for D to determine whether the generated pictures by G are true or not, so that D (G (z)) is 0.5, a model G of a generating equation is obtained, which can be used to generate pictures.
The objective function of the generative countermeasure network is:
Figure BDA0002958035770000081
the objective function indicates that for discriminant network D, it is desirable to maximize the objective function (discriminant formula V (D, G)), but for generating network G, it is desirable to minimize the objective function (discriminant formula V (D, G)), i.e., the entire training process is an iterative process, where D is the discriminant function, x is the true data, pdata (x) is the true data distribution, z is the generated data, pz (z) is the generated data distribution, E is the expected value, D (x) is the probability of discriminating the true data, and D (G (z)) is the probability of discriminating the generated data.
In an embodiment of the infrared identification method for LF furnace steel slag, before training the SegNet network model with the training data set, the method may further include building a SegNet network model:
establishing an input layer, wherein the input layer inputs slices of the infrared steel slag detection picture, specifically, the size of the picture can be any, and the picture can be preferably an infrared steel slag detection picture with 600 × 600 pixels in the embodiment;
building convolution layers, performing feature extraction and feature mapping by using convolution kernels, and extracting image features by using 4 groups of convolution layers;
it should be noted that, referring to fig. 3, fig. 3 is a schematic diagram of an overall architecture of an improved SegNet model, where conv (including conv1, conv2, conv3 and conv4) is a convolutional layer for extracting image features; the pool (including pool1, pool2, pool3 and pool4) is a pooling layer used for reducing dimension, removing redundant information, compressing features, simplifying network complexity, reducing calculation amount and memory consumption, namely reducing parameter amount because the parameter amount of deep learning is too large; deconv (including deconv1, deconv2, deconv3 and deconv4) is an deconvolution layer for reconstructing an image, unsamp (including unsamp1, unsamp2, unsamp3 and unsamp4) is an upsampling layer for filling the image in a copy manner, softmax is a normalized probability, which maps the output of a plurality of neurons into a (0,1) interval, which can be regarded as a probability, and classifies each pixel of the image as a multi-class; loss is the computational loss; label is used to classify and label each pixel in the image. In summary, the whole architecture of the SegNet model comprises an image input coding network which is composed of convolution layers, and the image input coding network extracts input features for target classification, maintains higher resolution, and reduces parameters; a decoding network, which is composed of deconvolution layers and maps the characteristic image with low resolution back to the same size as the input image; and (4) classifying the pixels, sending the decoded output to a classification layer, and finally generating class probability for each pixel independently to obtain an identification result.
The convolutional layer is improved on the basis of the original SegNet network convolutional layer, one group of convolution is removed, 4 groups of convolutional layers are used for extracting image characteristics, the convolutional layer utilizes convolution kernels for characteristic extraction and characteristic mapping, a picture with 600 pixels is input as an example, the first group of convolution groups adopts 7 pixels 64 convolution kernels, the convolution step length is 1, in order to generate output with the same size as the input picture, each grid pixel is expanded to the peripheral boundary of the original image, the expanded pixel value is 0, after the convolution kernel performs convolution operation on the image with the expanded boundary, 64 characteristic images are generated, and the value of the next layer of neurons generated by convolution kernel scanning is as follows:
Figure BDA0002958035770000091
in the formula, wijWeights for 7 x 64 convolution kernels, b bias terms for convolution kernels, wijAnd b may be trained, x being the number of regions covered by the convolution kernel. The computation of the convolutional layer is linear, so that an excitation layer is added behind the convolutional layer, and a nonlinear mapping is performed on the output of the convolutional layer, so that the network has nonlinear fitting capability, and the excitation function used in this embodiment is a ReLu function:
f(x)＝max(g(x),0) (3)
wherein g (x) represents the output of formula (2), and the function is to output the greater of the output value of formula 2 compared with 0, that is, if the value of formula 2 is negative, it will take 0. When an input image passes through a convolution layer, even if a small convolution kernel and step length are used, a large number of feature mapping maps can still be obtained, the calculation complexity is improved, the network training efficiency is reduced, at the moment, a down-sampling layer, namely a pooling layer can play a role, dimension reduction operation is carried out on each feature map, the operation amount of data is reduced, and meanwhile semantic abstraction can be carried out. The maximum pooling mode with kernel 2 x 2 and step 2 is used, that is, the maximum value of 4 "pooling visual field" matrixes is taken, the image size is reduced to 1/4 of the original image, the operation of the other 3 sets of convolution groups is consistent with that of the first set of convolution group, and finally the encoder outputs a feature map with size of 75 x 75.
And establishing a deconvolution layer, and performing up-sampling operation on the feature mapping image by using deconvolution to restore the image to the size of the original image.
Specifically, the establishing of the deconvolution layer is to implement end-to-end pixel level prediction of the infrared steel slag image, an up-sampling operation is performed on the feature mapping image by using deconvolution to restore the image to the size of an original image, the decoder network includes an inverse pooling operation and an inverse convolution operation, the embodiment uses the inverse pooling operation with a kernel of 4 and a step length of 2 to implement the up-sampling, one pixel is mapped into a plurality of pixels by recording position information of a label value in a corresponding encoder pooling layer, the following convolution operation is performed by using the inverse convolution kernel with the step length of 1 to enrich information of the sparse feature map, the operation of other 3 sets of the deconvolution layer sets is consistent with that of the first set of the deconvolution group, and finally, the size of the image output by the decoder is 600 × 600 pixels, which is the same as that of the original input image, thereby implementing end-to.
The above embodiment is described below with a specific example:
1. and constructing a data set, and intercepting the LF furnace molten steel infrared image from the field full-scene infrared image by the size of 600 × 600 pixels to serve as a data set image. 1350 images were taken and classified into a training data set and a test data set according to a 5:1 ratio.
2. Deep learning grid training: training is performed using a model of the SegNet network.
2.1 convolutional neural network architecture:
the deep learning multi-layer perceptron is specially designed for recognizing two-dimensional shapes and structurally mainly comprises a single or a plurality of convolution layers and a pooling layer. It has the ability to extract features more efficiently than other deep learning models due to its weight sharing and local perception.
In general, the convolution process is as equation
Figure BDA0002958035770000101
In this formula, l represents the number of layers, w represents the convolution kernel, MjRepresenting the selected feature map, and b is a weight.
2.2 structural parameters
Using the front end of the VGG-16 model as the encoder, the decoder is a mirror image of the encoder and has multiple upsampling, convolution, and activation functions, batch normalization is added to the output of each convolutional layer. For infrared steel slag detection, the decoder weights are initialized randomly and the entire network on the data set is fine tuned. Taking the two classification problem as an example, the ground-route detection graph of each image in the steel slag detection data set is converted into a 0-1 binary image. And when the defect foreground is separated from the general background, calculating a soft-max cross entropy loss function value, and adjusting the weight of the network back propagation to meet the minimum loss value. The cross entropy loss function is:
Figure BDA0002958035770000102
in this equation, lm is the label of pixel m in the image, qm is the probability that the pixel is steel slag, and the value of qm is obtained from the output of the network. The model is trained end-to-end using a small batch of stochastic gradient descent with momentum, learning rate, where learning rate is the step size. If the learning rate is too large, the algorithm will jump back and forth near the local optimal point and will not converge; if the learning rate is too small, the algorithm will move slowly in each step and the convergence rate will be slow, so a proper learning rate needs to be selected to achieve the result of convergence. The random gradient descent (SGD) algorithm employed in this embodiment only employs one sample to iterate each time, the training speed is fast, the result indicates that the loss value of the model is rapidly reduced before training, the prediction accuracy is also rapidly improved, and when the training is about 80 epochs, the model is stable.
3. Testing
Randomly inputting images which are not in the training set for testing, wherein the result is shown in fig. 4, fig. 4 is a schematic diagram of the test result of the infrared image of the molten steel in the test set, the left side is the original image, the right side is the result diagram, and the steel slag image can be accurately identified and segmented, and the area ratio is 32.29%.
In conclusion, by using the SegNet network model, more detailed characteristics of the steel slag image can be extracted, so that reliable detection of the steel slag is realized.
Fig. 5 shows an embodiment of an LF-furnace steel slag infrared identification system provided by the present invention, and fig. 5 is a schematic diagram of an embodiment of an LF-furnace steel slag infrared identification system provided by the present invention, where the system may include:
the image data acquisition device 501 is used for performing infrared detection on steel slag on the surface of a steel ladle of an LF furnace and bare molten steel in the steel ladle to obtain image data, and it needs to be explained that infrared thermal imaging is used as a non-contact detection means, so that the device has the advantages of rapidness, no damage, no contact, no need of coupling, real time, large detection range, intuition and the like, and can clearly observe an object to be monitored at night without light completely or in severe environments of waste gas and smoke dust of a steel mill. The infrared thermal imaging technology can improve the detection performance, can realize the quick, real-time and accurate detection of the steel ladle, prevent the interference of on-site light and dust, achieve the effect of stably obtaining the original image in the steel ladle, and transmit the obtained infrared image data to a subsequent preprocessing device;
the preprocessing device 502 is used for preprocessing the image data, and it should be noted that the preprocessing process may include, but is not limited to, image enhancement and noise reduction. Specifically, the method includes the steps of firstly calculating a multi-contrast combined image of absolute contrast, change contrast, normalized contrast, standard contrast and difference absolute contrast of an acquired infrared image to enhance a molten steel image in the infrared image, then denoising by using a wavelet transform threshold to remove uneven heating factors, noise and deformity in the image, wherein the infrared imaging system is easily interfered and uneven heating exists in the image Improper reconstruction methods or noise and malformations introduced by the external environment;
a data augmentation device 503, configured to perform data augmentation on the preprocessed image data to obtain a training data set and a test data set, where it should be noted that this data augmentation process is mainly used to add the training data set, so as to make the data set as diverse as possible, so that the trained model has a stronger generalization capability, including horizontal, vertical flipping, rotation, scaling, clipping, shearing, translation, contrast, color dithering, noise, and the like, and after the data augmentation, a large amount of image data is obtained, and then these data are divided into two parts, one part is used to train a SegNet network model, and the other part is used to test whether the model result has reached an expected standard after the model is trained, that is, whether the result is accurate;
the training device 504 is configured to train a SegNet Network model by using a training data set to obtain a deep learning model, and it is to be noted that the SegNet Network model mainly comprises an encoding Network (Encoder Network), a decoding Network (Decoder Network), and a Pixel-by-Pixel classifier (Pixel-by-Pixel Classification Layer), and each convolution Layer is followed by a Batch Normalization (Batch Normalization) Layer and a return activation function, where Batch Normalization of data is performed to make the data satisfy a normal distribution with a mean value of 0 and a variance of 1, and the Batch Normalization mainly functions to alleviate a gradient disappearance/explosion phenomenon in DNN training and accelerate a training speed of the model, the encoding Network converts a high-dimensional vector into a low-dimensional vector to implement low-dimensional extraction of high-dimensional features, and the encoding Network can capture more translational invariance features through multiple maximum pooling operations, but also can lose important basis such as boundary information of more feature maps, therefore, the maximum pooling index information is recorded simultaneously in the pooling process, the position of the maximum characteristic value is saved, then the input characteristic graph is up-sampled by using the maximum pooling index information, so that the boundary information is saved, the decoding network maps the characteristic graph with low resolution to the characteristic graph with high spatial resolution by using the maximum pooling index information of the corresponding characteristic layer saved in the down-sampling process of the encoder, the reconstruction from the low-dimensional vector to the high-dimensional vector is realized, the repeated use of the maximum pooling index in the decoding process has the advantages of optimizing the boundary profile description, reducing the parameter number and being capable of end-to-end training, the up-sampling mode can be applied to any encoding-decoding network, the high-dimensional characteristic expression vector is output by a decoder at the last layer and is used as the input of a trainable Softmax classifier, the SegNet network model is a full convolution network capable of pixel-level image segmentation, the infrared steel slag detection problem can be reconstructed into an end-to-end multi-classification problem;
the steel slag recognition device 505 is configured to test a test data set by using a deep learning model, and recognize steel slag in image data, and it should be noted that steel slag pixel points and non-steel slag pixel points are classified pixel by using strong learning ability of a SegNet deep learning network, so that a limitation that a traditional steel slag image detection manually extracts features and an algorithm is designed according to a specific problem by means of human experience is broken. The method is simple to operate, strong in robustness and high in accuracy, meets the real-time requirement of industrial steel slag image detection, and through experiments, the infrared steel slag detection accuracy can reach 98.5% by adopting a SegNet-based deep learning method.
The system can be used for quickly and accurately adjusting the flow of argon blown from the bottom of molten steel, has strong anti-interference capability, ensures the impurities in the molten steel to fully float, and stably controls the cleanliness of the molten steel.
In a specific embodiment of the above LF furnace steel slag infrared identification system, the image data acquisition device may be an infrared thermal imager installed above the LF furnace, and it should be noted that the infrared thermal imager receives an infrared radiation energy distribution image of the target to be detected by using an infrared detector and an optical imaging objective lens and reflects the infrared radiation energy distribution image onto a photosensitive element of the infrared detector, so as to obtain an infrared thermal image, and the thermal image corresponds to a thermal distribution field on the surface of the object. Infrared thermal imagers convert the invisible infrared energy emitted by an object into visible thermal images, the different colors on the thermal images representing the different temperatures of the object being measured. The non-contact infrared detection technology can be used for rapidly, accurately, conveniently and visually displaying the distribution of the surface temperature field of the object to be measured, measuring the surface temperature of the object, rapidly measuring the surface temperature reading of the object without directly contacting the surface of the object to be measured, reliably measuring the surface temperature of the object which is hot, dangerous or difficult to contact, and enabling the thermal infrared imager to be very fast in measurement speed and visually and continuously measure the temperature change of the surface of the object. Of course, this is only one preferred embodiment, and other types of infrared detection devices may be used according to actual needs, and are not limited herein.
In another specific embodiment of the LF furnace steel slag infrared identification system, the data amplification device is specifically configured to perform data amplification on the preprocessed image data through the generative countermeasure network at a small sample size. Specifically, in the machine learning and training process, the problem that the generalization performance (general-purpose capability and anti-interference capability) of the trained model is poor due to insufficient data volume (i.e., the samples are not comprehensive enough) is often encountered, at this time, the technology of artificially expanding the data to generate more kinds of data (more in line with the actual situation) is data expansion, and the acquired LF furnace molten steel infrared image data is subjected to data expansion through a Generative Adaptive Network (GAN) to generate more infrared images, so that the infrared image data set of the LF furnace molten steel is expanded, and thus a large number of data samples required by convolutional neural network training are obtained, and the problem that the convolutional neural network is difficult to train by a small sample data set is solved.
In a preferred embodiment of the LF furnace steel slag infrared identification system, the LF furnace steel slag infrared identification system may further include a SegNet network model building device, and may specifically include an input layer building unit, configured to build an input layer, where the input layer inputs a slice of an infrared steel slag detection picture, specifically, the size of the picture may be arbitrary, and the LF furnace steel slag infrared identification system may preferably be an infrared steel slag detection picture with 600 × 600 pixels in this embodiment;
the convolutional layer establishing unit is used for establishing a convolutional layer, performing feature extraction and feature mapping by using the convolutional layer, and extracting image features by using 4 groups of convolutional layers;
and the deconvolution layer establishing unit is used for establishing a deconvolution layer and performing up-sampling operation on the feature mapping image by using deconvolution so as to restore the image to the size of the original image.
Specifically, the establishing of the deconvolution layer is to implement end-to-end pixel level prediction of the infrared steel slag image, an up-sampling operation is performed on the feature mapping image by using deconvolution to restore the image to the size of an original image, the decoder network includes an inverse pooling operation and an inverse convolution operation, the embodiment uses the inverse pooling operation with a kernel of 4 and a step length of 2 to implement the up-sampling, one pixel is mapped into a plurality of pixels by recording position information of a label value in a corresponding encoder pooling layer, the following convolution operation is performed by using the inverse convolution kernel with the step length of 1 to enrich information of the sparse feature map, the operation of other 3 sets of the deconvolution layer sets is consistent with that of the first set of the deconvolution group, and finally, the size of the image output by the decoder is 600 × 600 pixels, which is the same as that of the original input image, thereby implementing end-to.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (10)

1. An LF furnace steel slag infrared identification method is characterized by comprising the following steps:
carrying out infrared detection on steel slag on the surface of a steel ladle of the LF furnace and bare molten steel in the steel ladle to obtain image data;
preprocessing the image data;
performing data augmentation on the preprocessed image data to obtain a training data set and a test data set;
training a SegNet network model by using the training data set to obtain a deep learning model;
and testing the test data set by using the deep learning model, and identifying the steel slag in the image data.
2. The LF furnace steel slag infrared identification method as claimed in claim 1, wherein an infrared thermal imager arranged above the LF furnace is used for performing infrared detection on steel slag on the surface of a steel ladle of the LF furnace and bare molten steel in the steel ladle.
3. The LF furnace steel slag infrared identification method as claimed in claim 1, wherein the preprocessing the image data comprises:
and performing enhancement and noise reduction processing on the image data.
4. The LF furnace steel slag infrared identification method as claimed in claim 1, wherein the data augmentation of the preprocessed image data comprises:
and carrying out data augmentation on the preprocessed image data under a small sample amount through a generative countermeasure network.
5. The LF furnace steel slag infrared identification method as claimed in claim 1, wherein before training the SegNet network model by using the training data set, the method further comprises the following steps of:
establishing an input layer, wherein the input layer inputs the slices of the infrared steel slag detection picture;
building convolution layers, performing feature extraction and feature mapping by using convolution kernels, and extracting image features by using 4 groups of convolution layers;
and establishing a deconvolution layer, and performing up-sampling operation on the feature mapping image by using deconvolution to restore the image to the size of the original image.
6. An LF stove slag infrared identification system which characterized in that includes:
the image data acquisition device is used for carrying out infrared detection on steel slag on the surface of the steel ladle of the LF furnace and bare molten steel in the steel ladle to obtain image data;
preprocessing means for preprocessing the image data;
the data amplification device is used for performing data amplification on the preprocessed image data to obtain a training data set and a test data set;
the training device is used for training the SegNet network model by utilizing the training data set to obtain a deep learning model;
and the steel slag identification device is used for testing the test data set by using the deep learning model and identifying the steel slag in the image data.
7. The LF furnace steel slag infrared identification system as claimed in claim 6, wherein the image data acquisition device is a thermal infrared imager mounted above the LF furnace.
8. The LF furnace steel slag infrared identification system as claimed in claim 6, wherein the preprocessing device is specifically configured to perform enhancement and noise reduction processing on the image data.
9. The LF furnace steel slag infrared identification system as claimed in claim 6, wherein the data amplification device is specifically configured to perform data amplification on the preprocessed image data through a generative countermeasure network under a small sample amount.
10. The LF furnace steel slag infrared identification system as claimed in claim 6, further comprising a SegNet network model building device, specifically comprising an input layer building unit for building an input layer, wherein the input layer inputs the slices of the infrared steel slag detection pictures;
the convolutional layer establishing unit is used for establishing a convolutional layer, performing feature extraction and feature mapping by using the convolutional layer, and extracting image features by using 4 groups of convolutional layers;
and the deconvolution layer establishing unit is used for establishing a deconvolution layer and performing up-sampling operation on the feature mapping image by using deconvolution so as to restore the image to the size of the original image.
CN202110228878.0A 2021-03-02 2021-03-02 LF furnace steel slag infrared identification method and system Pending CN112950586A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110228878.0A CN112950586A (en) 2021-03-02 2021-03-02 LF furnace steel slag infrared identification method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110228878.0A CN112950586A (en) 2021-03-02 2021-03-02 LF furnace steel slag infrared identification method and system

Publications (1)

Publication Number Publication Date
CN112950586A true CN112950586A (en) 2021-06-11

Family

ID=76247102

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110228878.0A Pending CN112950586A (en) 2021-03-02 2021-03-02 LF furnace steel slag infrared identification method and system

Country Status (1)

Country Link
CN (1) CN112950586A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113714496A (en) * 2021-07-20 2021-11-30 武汉钢铁有限公司 Ladle bottom argon blowing fault diagnosis method and device and storage medium
CN115679042A (en) * 2022-11-04 2023-02-03 唐山惠唐物联科技有限公司 Method and system for monitoring smelting state in refining process of LF (ladle furnace)
CN116168348A (en) * 2023-04-21 2023-05-26 成都睿瞳科技有限责任公司 Security monitoring method, system and storage medium based on image processing
CN116339128A (en) * 2023-05-30 2023-06-27 北京国电富通科技发展有限责任公司 Online monitoring method and system for slag falling rainfall of slag conveying system

Citations (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040187641A1 (en) * 2000-03-17 2004-09-30 Kemeny Frank L. Slag detector for molten steel transfer operations
CN103571994A (en) * 2012-08-01 2014-02-12 宝山钢铁股份有限公司 Infrared steel slag detection method of converter
CN205188338U (en) * 2015-12-07 2016-04-27 攀钢集团西昌钢钒有限公司 Converter tapping is sediment control system down
US20180172601A1 (en) * 2016-12-15 2018-06-21 Thyssenkrupp Rasselstein Gmbh Method of inspecting a steel strip
CN108986098A (en) * 2018-09-05 2018-12-11 中冶赛迪技术研究中心有限公司 A kind of molten iron intelligence slag skimming method based on machine vision
CN110413013A (en) * 2019-07-18 2019-11-05 莱芜钢铁集团电子有限公司 A kind of intelligence argon blowing system and its control method
CN110438284A (en) * 2019-08-26 2019-11-12 杭州谱诚泰迪实业有限公司 A kind of converter intelligence tapping set and control method
CN110532902A (en) * 2019-08-12 2019-12-03 北京科技大学 A kind of molten iron drossing detection method based on lightweight convolutional neural networks
CN110796046A (en) * 2019-10-17 2020-02-14 武汉科技大学 Intelligent steel slag detection method and system based on convolutional neural network
CN111028253A (en) * 2019-11-25 2020-04-17 北京科技大学 Iron concentrate powder segmentation method and segmentation device
CN111679622A (en) * 2019-03-11 2020-09-18 上海梅山钢铁股份有限公司 Device and method for regulating and controlling argon flow of bottom blowing of steel ladle
CN112017145A (en) * 2019-05-31 2020-12-01 宝山钢铁股份有限公司 Efficient molten iron pretreatment automatic slag skimming method and system

Patent Citations (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040187641A1 (en) * 2000-03-17 2004-09-30 Kemeny Frank L. Slag detector for molten steel transfer operations
CN103571994A (en) * 2012-08-01 2014-02-12 宝山钢铁股份有限公司 Infrared steel slag detection method of converter
CN205188338U (en) * 2015-12-07 2016-04-27 攀钢集团西昌钢钒有限公司 Converter tapping is sediment control system down
US20180172601A1 (en) * 2016-12-15 2018-06-21 Thyssenkrupp Rasselstein Gmbh Method of inspecting a steel strip
CN108986098A (en) * 2018-09-05 2018-12-11 中冶赛迪技术研究中心有限公司 A kind of molten iron intelligence slag skimming method based on machine vision
CN111679622A (en) * 2019-03-11 2020-09-18 上海梅山钢铁股份有限公司 Device and method for regulating and controlling argon flow of bottom blowing of steel ladle
CN112017145A (en) * 2019-05-31 2020-12-01 宝山钢铁股份有限公司 Efficient molten iron pretreatment automatic slag skimming method and system
CN110413013A (en) * 2019-07-18 2019-11-05 莱芜钢铁集团电子有限公司 A kind of intelligence argon blowing system and its control method
CN110532902A (en) * 2019-08-12 2019-12-03 北京科技大学 A kind of molten iron drossing detection method based on lightweight convolutional neural networks
CN110438284A (en) * 2019-08-26 2019-11-12 杭州谱诚泰迪实业有限公司 A kind of converter intelligence tapping set and control method
CN110796046A (en) * 2019-10-17 2020-02-14 武汉科技大学 Intelligent steel slag detection method and system based on convolutional neural network
CN111028253A (en) * 2019-11-25 2020-04-17 北京科技大学 Iron concentrate powder segmentation method and segmentation device

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
R. STRĄKOWSKI 等: "Estimation of FeO content in the steel slag using infrared imaging and artificial neural network", 《MATERIALS SCIENCE》 *
RUBÉN USAMENTIAGA 等: "Temperature Measurement of Molten Pig Iron With Slag Characterization and Detection Using Infrared Computer Vision", 《IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT》 *
焦李成 等: "《人工智能前沿技术丛书 计算智能导论》", 30 September 2019, 西安电子科技大学出版社 *
王卫东 等: "基于深度学习的煤中异物机器视觉检测", 《矿业科学学报》 *

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113714496A (en) * 2021-07-20 2021-11-30 武汉钢铁有限公司 Ladle bottom argon blowing fault diagnosis method and device and storage medium
CN113714496B (en) * 2021-07-20 2023-01-24 武汉钢铁有限公司 Ladle bottom argon blowing fault diagnosis method and device and storage medium
CN115679042A (en) * 2022-11-04 2023-02-03 唐山惠唐物联科技有限公司 Method and system for monitoring smelting state in refining process of LF (ladle furnace)
CN116168348A (en) * 2023-04-21 2023-05-26 成都睿瞳科技有限责任公司 Security monitoring method, system and storage medium based on image processing
CN116168348B (en) * 2023-04-21 2024-01-30 成都睿瞳科技有限责任公司 Security monitoring method, system and storage medium based on image processing
CN116339128A (en) * 2023-05-30 2023-06-27 北京国电富通科技发展有限责任公司 Online monitoring method and system for slag falling rainfall of slag conveying system
CN116339128B (en) * 2023-05-30 2023-07-28 北京国电富通科技发展有限责任公司 Online monitoring method and system for slag falling rainfall of slag conveying system

Similar Documents

Publication Publication Date Title
CN112950586A (en) LF furnace steel slag infrared identification method and system
CN108615226B (en) Image defogging method based on generation type countermeasure network
CN112861635B (en) Fire disaster and smoke real-time detection method based on deep learning
CN112232349A (en) Model training method, image segmentation method and device
CN109948566B (en) Double-flow face anti-fraud detection method based on weight fusion and feature selection
CN111986142A (en) Unsupervised enhancement method for surface defect image data of hot-rolled plate coil
Liu et al. A night pavement crack detection method based on image‐to‐image translation
CN109961042B (en) Smoke detection method combining deep convolutional neural network and visual change diagram
CN111539343A (en) Black smoke vehicle detection method based on convolution attention network
CN112926652B (en) Fish fine granularity image recognition method based on deep learning
CN109815653A (en) A kind of extraction of pdf Text Watermarking and comparison method based on deep learning
CN109598200B (en) Intelligent image identification system and method for molten iron tank number
CN113553954A (en) Method and apparatus for training behavior recognition model, device, medium, and program product
CN112766056A (en) Method and device for detecting lane line in low-light environment based on deep neural network
CN112767344A (en) Disease enhancement method based on vehicle-mounted camera shooting and coupling tradition and machine learning
CN114359733A (en) Vision-based smoke fire detection method and system
CN115861210A (en) Transformer substation equipment abnormity detection method and system based on twin network
CN116030361A (en) CIM-T architecture-based high-resolution image change detection method
CN114972246A (en) Die-cutting product surface defect detection method based on deep learning
CN114926826A (en) Scene text detection system
CN112016437B (en) Living body detection method based on face video key frame
CN112270404A (en) Detection structure and method for bulge defect of fastener product based on ResNet64 network
CN115953312A (en) Joint defogging detection method and device based on single image and storage medium
CN113192018B (en) Water-cooled wall surface defect video identification method based on fast segmentation convolutional neural network
CN111402223B (en) Transformer substation defect problem detection method using transformer substation video image

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20210611

RJ01 Rejection of invention patent application after publication