CN113537226A - Smoke detection method based on deep learning - Google Patents
Smoke detection method based on deep learning Download PDFInfo
- Publication number
- CN113537226A CN113537226A CN202110542645.8A CN202110542645A CN113537226A CN 113537226 A CN113537226 A CN 113537226A CN 202110542645 A CN202110542645 A CN 202110542645A CN 113537226 A CN113537226 A CN 113537226A
- Authority
- CN
- China
- Prior art keywords
- smoke
- gaussian
- detection
- yolov3
- matching
- 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
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a smoke detection method based on deep learning, and relates to the technical field of smoke detection; the detection method comprises the following steps: step one, fusing an improved mixed Gaussian with YOLOv3 to a smoke detection algorithm; 1.1, improving the foreground extraction of the mixed Gaussian smoke; 1.2, YOLOv3 model and training; step two, a smoke detection algorithm based on improved YOLOv 3; the smoke detection method utilizes the improved Gaussian mixture to extract the smoke foreground, eliminates static interference and simultaneously reduces the smoke detection range; then extracting by adopting a target detection algorithm, eliminating static interference and simultaneously reducing the range of smoke detection; then YOLOv3 trains data and extracts smoke features; finally, fusing the two, predicting the specific position and range of the smoke, framing the smoke and completing the detection of the smoke; the smoke detector can realize rapid and accurate detection of smoke in a complex scene.
Description
Technical Field
The invention belongs to the technical field of smoke detection, and particularly relates to a smoke detection method based on deep learning.
Background
The fire disaster is a disaster caused by improper fire utilization or spontaneous combustion in the nature, if the fire is not controlled in time, the fire disaster lasts for a long time, threatens lives in a large-scale area, and has irreversible influence on the natural environment. Among various natural disasters, a fire disaster is one of the most common disasters threatening public safety and social development most frequently, smoke is generated and scattered continuously in the early stage of the fire disaster, and if the smoke can be detected in time and effectively suppressed in the early stage, property loss and social influence can be minimized, so that the smoke detection in the early stage of the fire disaster is very important. Most of the traditional smoke detection methods are based on detection technologies of physical sensors, including ion-type smoke sensors, photoelectric smoke sensors, etc., but are not suitable for environments with high sensitivity requirements or complex, diverse and wide ranges. The traditional smoke detection device is small in detection range, high in equipment cost, extremely easy to be influenced by weather illumination and the like, and high in false smoke alarm rate.
At present, most of the devices for fire detection and alarm are smoke sensors, which are generally installed in indoor places where fire easily occurs, such as parking lots, factories, enterprise office areas, homes and other environments, and are widely used due to low production cost and simple installation. The detector triggers the physical element to sense the fire according to the conditions of smoke concentration, ambient temperature, change of ambient brightness after the fire becomes large and the like generated when the fire occurs. However, it has a great limitation in that a fire can be detected only when the smoke concentration or the ambient temperature rises to a certain level, and thus it is generally applied only in a small space, and its sensitivity is lowered with the passage of time, so that the fire, concentration and other influences cannot be detected even in a small space, and the sensor signal has a slow transmission rate and becomes weak and ineffective, so that the recognition effect is poor, the false alarm rate is high, and the optimal rescue and rescue time is lost. The device is mainly composed of two research branches, one is mainly used for flame detection, and the other is mainly used for smoke detection.
From the practical application, the technology mainly based on smoke detection is more advantageous because the flame is generally not obvious in the initial stage of fire occurrence, especially in the high and dense zones such as forest, the flame is easily shielded, and the smoke generated by fire is easily captured by the camera due to the property that the smoke floats upwards and the area of the smoke gradually increases. If the smoke can be detected in time, measures can be taken in advance to control the fire. With the development of image recognition technology and the wide application of video monitoring facilities, the smoke recognition method based on video becomes a research hotspot of scholars at home and abroad. However, the smoke form is not fixed and can be diffused randomly, more targets like smoke exist in a smoke field, the positioning and the detection of real smoke are not facilitated, and the false alarm phenomenon is high. Therefore, the accuracy of smoke video detection is continuously improved, the false alarm rate and the missing alarm rate are reduced, and the method has high practical value for timely and effectively discovering and inhibiting the fire. The video-based smoke identification method becomes a research hotspot of scholars at home and abroad.
Disclosure of Invention
The method aims to solve the problems that the existing smog is unfixed in shape and randomly diffused, more smog-like targets exist in a smog site, the positioning and the detection of real smog are not facilitated, and the false alarm phenomenon is high; the invention aims to provide a smoke detection method based on deep learning.
The invention discloses a smoke detection method based on deep learning, which comprises the following steps:
step one, an improved mixed Gaussian and YOLOv3 fused smoke detection algorithm:
is divided into two parts; the first part carries out pixel matching on each frame of a video by utilizing an improved Gaussian mixture algorithm to obtain a corresponding static background and a corresponding dynamic foreground, and then extracts the dynamic foreground according to a certain rule to obtain a coarsely screened smoke foreground; the second part mainly adopts a YOLOv3 target detection algorithm to extract smoke characteristics and carry out secondary fine screening; when the smoke features are extracted by using YOLOv3, the preprocessed smoke data set is used as the input of a model, repeated training and parameter adjustment are carried out in the model, finally, the weight for representing the smoke features is generated, calling is carried out in the smoke foreground obtained by rough screening, secondary fine screening is carried out on the smoke region, and the specific position and range of the video smoke are determined finally, so that the smoke is effectively detected;
1.1, improving mixed Gaussian smoke foreground extraction:
1.1.1) matching pixel points of a current video frame with Gaussian distribution, if the formula (1-2) is satisfied, matching the pixel points with the Gaussian distribution, and if the formula is not satisfied, indicating that the pixel points are not matched;
|Xi,t-μi|＜2.5σi (1-2)
1.1.2) if the matching is successful, updating the Gaussian background model by using the pixel point, wherein the updated formula is represented by the formula (1-3), the formula (1-4) and the formula (1-5):
ωi,t＝(1-α)ωi-1+α(Mi,t) (1-3)
μi,t＝(1-θ)μi,t-1+θXt (1-4)
σ2 i,t＝(1-θ)σ2 i,t-1+θ(Xt-μi,t-1)T(Xt-μi,t-1) (1-5)
wherein α is a learning rate; mi,tThe weight change of the multiple Gaussian distributions is controlled. When the matching is successful, Mi,t1 is ═ 1; when the matching fails, Mi,t0; alpha and theta may reflect the ability of gaussians to adapt to different scenarios;
1.1.3) if the distribution does not match any Gaussian distribution, adding a Gaussian distribution or updating the Gaussian distribution to replace the distribution with the minimum priority;
the K Gaussian distributions are arranged from high to low according to the priority, and are represented by the following formula (1-6):
1.1.4) after updating parameters of the Gaussian model, arranging K Gaussian distributions of pixel points according to priority, and taking the first B Gaussian distributions as ideal description of background pixels:
in the formula, ωkRepresenting the weight of the Kth Gaussian function in one pixel; t is typically taken to be between 0.5 and 1, where an empirical value of 0.85 is taken, and continuing on for Xi,tAnd performing matching test with the B Gaussian distributions, wherein the matching with any one of the first B Gaussian distributions is the foreground, and the matching is the background if the matching is not the foreground.
1.2, YOLOv3 model and training:
predicting a target area and a target category in a neural network model; YOLOv3 adopts a new Darknet-53 network structure of ResNet idea, adds a residual error module in the network, and adopts 3 feature maps with different scales to detect objects, so that more fine-grained features can be detected;
step two, a smoke detection algorithm based on the improved YOLOv 3:
the channel attention mechanism is nested in the last layer of a Residual convolution Block Residul Block in a YOLOv3 network structure, in smoke target detection, when two or more target frames are close, the frame with lower score is deleted because of overlarge overlapping area, the threshold value of NMS needs to be manually determined, detection is missed when the threshold value is small, false detection is carried out when the threshold value is large, and non-maximum value inhibition is changed into Soft-NMS.
Preferably, the Soft-NMS firstly generates some smoke candidate areas in smoke detection, secondly obtains category scores through a classification network, meanwhile obtains more accurate position parameters through a regression network, and finally obtains the final detection result through non-maximum suppression.
Compared with the prior art, the invention has the beneficial effects that:
firstly, extracting a smoke foreground by using an improved Gaussian mixture, and reducing the range of smoke detection while eliminating static interference; then extracting by adopting a target detection algorithm, eliminating static interference and simultaneously reducing the range of smoke detection; then YOLOv3 trains data and extracts smoke features; and finally fusing the two, predicting the specific position and range of the smoke, framing the smoke and completing the detection of the smoke.
Secondly, the rapid and accurate detection of smoke in a complex scene can be realized.
Drawings
For ease of illustration, the invention is described in detail by the following detailed description and the accompanying drawings.
FIG. 1 is a diagram of a dynamic foreground region in the present invention;
FIG. 2 is a flow chart of a fusion algorithm in the present invention;
FIG. 3 is a schematic diagram of the YOLOv3 algorithm structure in the present invention;
fig. 4 is a schematic diagram of an ECA module of the invention.
Detailed Description
In order that the objects, aspects and advantages of the invention will become more apparent, the invention will be described by way of example only, and in connection with the accompanying drawings. It is to be understood that such description is merely illustrative and not intended to limit the scope of the present invention. The structure, proportion, size and the like shown in the drawings are only used for matching with the content disclosed in the specification, so that the person skilled in the art can understand and read the description, and the description is not used for limiting the limit condition of the implementation of the invention, so the method has no technical essence, and any structural modification, proportion relation change or size adjustment still falls within the range covered by the technical content disclosed by the invention without affecting the effect and the achievable purpose of the invention. Moreover, in the following description, descriptions of well-known structures and techniques are omitted so as to not unnecessarily obscure the concepts of the present invention.
It should be noted that, in order to avoid obscuring the present invention with unnecessary details, only the structures and/or processing steps closely related to the scheme according to the present invention are shown in the drawings, and other details not so relevant to the present invention are omitted.
The specific implementation mode adopts the following technical scheme:
an improved mixed Gaussian and YOLOv3 fused smoke detection algorithm:
aiming at the problems that a suspected fire smoke area is difficult to screen, smoke scenes are greatly different, and smoke detection accuracy is low, the smoke detection method for improving the fusion of Gaussian mixture and YOLOv3 is provided. The fusion algorithm is mainly divided into 2 parts. And the part 1 carries out pixel matching on each frame of the video by utilizing an improved Gaussian mixture algorithm to obtain a corresponding static background and a corresponding dynamic foreground, and then extracts the dynamic foreground according to a certain rule to obtain a coarsely screened smoke foreground. In the part 2, a YOLOv3 target detection algorithm is mainly adopted to extract smoke features and carry out secondary fine screening. When the smoke features are extracted by using the YOLOv3, the preprocessed smoke data set is used as the input of a model, repeated training and parameter adjustment are carried out in the model, finally, the weight for representing the smoke features is generated, calling is carried out in the smoke foreground obtained by rough screening, secondary fine screening is carried out on the smoke region, and the specific position and range of the video smoke are determined finally, so that the smoke is effectively detected.
1.1, improving mixed Gaussian smoke foreground extraction:
the adaptive Gaussian mixture model proposed by Stauffer and the like uses the weighting and modeling of k Gaussian functions to extract the vector characteristics of all pixel points in a moving image, matches and updates the Gaussian mixture model data in the next frame, and separates the foreground from the background through the matching of the pixel points.
The weighted sum of the probability density functions of the current pixel points is as follows:
in the formula, K represents the number of Gaussian distribution and is generally between 3 and 5; omegai,tRepresenting the weight of the ith Gaussian function at the moment t; eta (X)i,μi,t,∑i,t) Is the ith Gaussian probability density function; mu.si,tRepresents the mean of the ith gaussian function at time t; sigmai,tRepresenting the variance of the ith gaussian function at time t. Assuming that R, G, B channel pixel values in the detected image are independent of each other and have the same variance, take Σi,t＝σ2 i,tI. Wherein the mean and the square of the Gaussian functionThe difference has a large effect on the overall mixed gaussian model.
The main steps of the mixed gauss extraction foreground are as follows:
1) and matching the pixel points of the current video frame with the Gaussian distribution, if the formula (1-2) is satisfied, matching the pixel points with the Gaussian distribution, and if the formula is not satisfied, indicating that the pixel points are not matched.
|Xi,t-μi|＜2.5σi (1-2)
2) If the matching is successful, updating the Gaussian background model by using the pixel point, wherein the updated formula is the following formula (1-3), the formula (1-4) and the formula (1-5):
ωi,t＝(1-α)ωi-1+α(Mi,t) (1-3)
μi,t＝(1-θ)μi,t-1+θXt (1-4)
σ2 i,t＝(1-θ)σ2 i,t-1+θ(Xt-μi,t-1)T(Xt-μi,t-1) (1-5)
wherein α is a learning rate; mi,tThe weight change of the multiple Gaussian distributions is controlled. When the matching is successful, Mi,t1 is ═ 1; when the matching fails, Mi,t0; alpha and theta may reflect the ability of gaussian to adapt to different scenarios.
3) If there is no match with either gaussian, then one gaussian (larger variance and smaller weight) is added or updated instead of the lowest priority distribution.
The K Gaussian distributions are arranged from high to low according to the priority, and are represented by the following formula (1-6):
4) after updating parameters of the Gaussian model, arranging K Gaussian distributions of the pixel points according to the priority, and taking the first B Gaussian distributions as ideal description of background pixels:
in the formula, ωkRepresenting the weight of the Kth Gaussian function in one pixel; t is typically taken to be between 0.5 and 1, where an empirical value of 0.85 is taken, and continuing on for Xi,tAnd performing matching test with the B Gaussian distributions, wherein the matching with any one of the first B Gaussian distributions is the foreground, and the matching is the background if the matching is not the foreground.
Compared with other moving object detection methods, such as an optical flow method and an interframe difference method, the foreground extraction method of the mixed gaussians has high adaptability to the background, can better describe a complex environment, and has a background elimination effect which depends on morphological characteristics of a video environment and an object to a certain extent. When a fire disaster happens, smoke is scattered, and is easily influenced by external factors such as the surrounding environment, the wind speed, the wind direction and the like to be randomly diffused, the average gradient of each pixel point is correspondingly changed along with the continuous accumulation or diffusion of the smoke, and the overall outline is complex and fuzzy.
And testing and comparing smoke videos of different scenes by integrating the characteristics of the smoke of the fire. With the diffusion of smoke and the continuation of time, the mixed Gaussian model only detects a small amount of peripheral parts of smoke diffusion, and the gradually-filled inner-layer smoke is processed as a background, so that the extraction of the video smoke foreground by using the pure mixed Gaussian is not ideal. And the diffusion form and the unique expansion ascending motion of the smoke are combined, and the Gaussian mixture model is improved, so that the preliminary screening of the suspected smoke area is realized.
Based on the inherent characteristics of smoke, the method improves the Gaussian mixture to adapt to the extraction of the smoke prospect, and comprises the following specific improvement steps:
1) the Gaussian mixture model takes the matched pixel points of each frame as a black foreground, takes the unmatched pixel points as a white background, and converts each frame corresponding to the video into a binary image.
2) Aiming at each matched frame of binary image, a coordinate system is established by taking the upper left corner of the image as an origin, all pixel points in the image are traversed, and the minimum value (x) corresponding to the x-axis coordinate and the y-axis coordinate is found from all black pixel pointsmin,ymin) And maximum value (x)max,ymax) Accordingly, the corresponding smoke foreground interested region is defined, as shown in fig. 1.
3) The delineated smoke foreground region of interest may be expressed as
[xmax+μ∶xmin-μ,ymax+β∶ymin-β]This is indicated by a dotted rectangle in FIG. 1 (b). When μ ═ 0 and β ═ 0, the smoke foreground region of interest is the middle solid line rectangular portion in fig. 1 (b); the values of the mu and the beta can be properly adjusted according to different video scenes to achieve a good smoke region extraction effect. And if no black pixel point exists in the matched binary image, namely no motion foreground is extracted, carrying out detection updating on the next frame.
4) And feeding back the coordinate information of the smoke foreground interesting region defined by each frame of binary image to the original image corresponding to the video, extracting the smoke region and finishing the first coarse screening. Through experiments on fire smoke in different scenes, the improved Gaussian mixture can effectively extract the smoke prospect.
1.2, YOLOv3 model and training:
based on the real-time requirement of smoke detection, a YOLOv3 target detection algorithm with strong real-time performance and better accuracy is selected from the smoke detection algorithm. The algorithm is a single-step detection algorithm, and target areas and target categories are predicted in a neural network model. Compared with the networks of two-step methods such as Faster R-CNN and Marsk R-CNN, the method has similar accuracy and has the advantage of instantaneity which is not possessed by the networks.
YOLOv3 adopts a new darknet-53 network structure of ResNet idea, adds a residual error module in the network, and adopts 3 feature maps with different scales to detect objects, thereby being capable of detecting features with finer granularity. The final output of the network is a profile of 1/32, 1/16, and 1/8 for 3 scales, respectively. After 79 th layer, 1/32(13 × 13) prediction results are obtained through several convolution operations, the down-sampling multiple is high, the perception field of the feature map is large, and therefore the feature map is suitable for detecting objects with large sizes in the image, the results are subjected to tensor stitching with the results of 61 st layer through up-sampling, the prediction results of 1/16 are obtained through several convolution operations and are suitable for detecting objects with medium sizes, the results of 61 st layer are subjected to tensor stitching with the results of 36 th layer after up-sampling, and the results are 1/8 after several convolution operations and are suitable for detecting objects with small sizes. Finally, we predict the object class with respect to YOLOv3 without using softmax and instead using the output of logistic. Such operations can support multi-tagged objects.
By carrying out data acquisition and detection experiments on common smoke scenes, the smoke video is segmented into frames, calibrated and made into a smoke data set. The fusion algorithm flow chart is shown in fig. 2.
Second, smoke detection algorithm based on improved YOLOv 3:
as shown in fig. 3, in order to obtain higher accuracy, the complexity of the model and the computational load can be reduced. A channel attention mechanism (ECA-Net) is nested in the last layer of a Residual volume Block in a YOLOv3 network structure, dimension reduction is avoided, and cross-channel interaction information is effectively captured. In the smoke target detection, when two or more target frames are close to each other, the frame with lower score is deleted because the overlapping area of the frame with lower score is too large, the threshold value of the NMS needs to be manually determined, the frame is set to be small, the frame is set to be missed, the frame is set to be false, and the non-maximum suppression (NMS) is changed into Soft-NMS. Through the improvement, the improved YOLOv3 smoke detection algorithm realizes the rapid and accurate detection of smoke in a complex scene.
2.1, channel attention mechanism (ECA-Net):
the deep convolutional neural network is widely applied to the field of computer vision, and makes great progress in the fields of image classification, target detection, semantic segmentation and the like. Starting from the pioneering AlexNet, new CNN models are continuously being derived in order to further improve the performance of deep convolutional neural networks. In recent years, the introduction of channel attention to volume blocks has attracted a great deal of attention, showing great potential in performance improvement. The representative method is SE-Net, which can learn the channel attention of each volume block, resulting in significant performance gains for various deep CNN architectures. Although these methods achieve higher accuracy, they often bring higher model complexity and larger computational burden. Therefore, an Efficient Channel Attention (ECA) module for a deep convolutional neural network is provided, which avoids dimension reduction and effectively captures cross-channel interaction information. As shown in fig. 4.
After channel-wise global average pooling without dimensionality reduction, the ECA captures local cross-channel mutual information by considering each channel and its k neighbors. Practice proves that the method ensures the model efficiency and the calculation effect. It is noted that ECA can be efficiently implemented by fast one-dimensional convolution with a size k, where the convolution kernel size k represents the coverage of local cross-channel interactions, i.e. how many neighbors around the channel participate in the attention prediction of this channel, and to avoid manual tuning of k by cross-validation, a method is proposed to adaptively determine k, where the coverage of interactions (i.e. the convolution kernel size k) is proportional to the channel dimension.
Compared to the backbone model, the deep convolutional neural network with ECA modules introduces few extra parameters and almost negligible computations. The modules are formed by nonlinear adaptively determined one-dimensional convolutions. The ECA is a very lightweight plug-and-play block that can leverage the performance of various CNNs. The ECA module learns effective channel attention by avoiding reducing channel dimensions while acquiring cross-channel interaction information in an extremely lightweight manner.
Therefore, a channel attention mechanism (ECA-Net) is nested in the last layer of a Residual Block in a YOLOv3 network structure, so that the complexity and the computational burden of a model can be reduced on the premise of obtaining higher precision.
2.2、Soft-NMS：
In the smoke detection, firstly, some smoke candidate areas are generated, secondly, category scores are obtained through a classification network, meanwhile, more accurate position parameters are obtained through a regression network, and finally, the final detection result is obtained through non-maximum suppression.
The smoke detection algorithm may output multiple detection frames, especially there may be many detection frames with high confidence around the real target. In order to remove repeated detection frames, the purpose that each detection object has only one detection result is achieved. Non-maximum suppression (NMS), an algorithm for obtaining local maxima and suppressing non-maxima, has wide application in computer vision. Specific steps of non-maximum suppression: setting an IOU threshold value for each category, sorting the smoke candidate frames according to the category scores, selecting the smoke candidate frame with the highest category score, traversing the rest candidate frames, if the IOU of the smoke candidate frame with the highest category score with the current category score is larger than the IOU threshold value, removing the smoke candidate frame, continuously selecting one smoke frame with the highest category score from the unprocessed smoke frames, and repeating the process until all frames are processed. And the reserved smoke candidate frame is the detection result. The conventional non-maximum suppression algorithm first generates a series of detection boxes B and corresponding scores S in the detected picture. When the largest scoring test box M is selected, it is removed from set B and placed in the final test result set D. At the same time, any detection frame in the set B that overlaps the detection frame M by more than the overlap threshold will be removed. If an object appears in the overlapping area of another object, that is, when two target frames are close to each other, the lower-score frame is deleted because the overlapping area is too large, thereby causing the detection of the object to fail and reducing the average detection rate of the algorithm, and the algorithm is not reasonable to design. The non-maximum suppression (NMS) is changed to Soft-NMS. The Soft-NMS algorithm does not directly delete all boxes whose IOU is greater than the threshold but reduces its confidence.
Nesting the last layer of Residual Block in the YOLOv3 network structure through a channel attention mechanism (ECA-Net), and changing the non-maximum suppression (NMS) to Soft-NMS. Through the improvement, the smoke can be rapidly and accurately detected in a complex scene.
The effect of the embodiment is as follows:
1. in the fusion algorithm, firstly, the smoke foreground is extracted by using an improved Gaussian mixture, static interference is eliminated, and the smoke detection range is narrowed; then extracting by adopting a target detection algorithm, eliminating static interference and simultaneously reducing the range of smoke detection; then YOLOv3 trains data and extracts smoke features; and finally fusing the two, predicting the specific position and range of the smoke, framing the smoke and completing the detection of the smoke.
2. A channel attention mechanism (ECA-Net) is nested in the last layer of a Residual Block in a YOLOv3 network structure, the model avoids dimension reduction, reduces the calculation load, effectively captures cross-channel interaction information, and improves the rapid detection of smoke. And the non-maximum value suppression (NMS) is changed into Soft-NMS, so that the condition of missing detection of smoke detection is reduced, and the accuracy of smoke detection is improved. Therefore, the smoke can be rapidly and accurately detected in a complex scene.
It will be evident to those skilled in the art that the invention is not limited to the details of the foregoing illustrative embodiments, and that the present invention may be embodied in other specific forms without departing from the spirit or essential attributes thereof. The present embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of the invention being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein.
Furthermore, it should be understood that although the present description refers to embodiments, not every embodiment may contain only a single embodiment, and such description is for clarity only, and those skilled in the art should integrate the description, and the embodiments may be combined as appropriate to form other embodiments understood by those skilled in the art.
Claims (2)
1. A smoke detection method based on deep learning is characterized in that: the detection method comprises the following steps:
step one, an improved mixed Gaussian and YOLOv3 fused smoke detection algorithm:
is divided into two parts; the first part carries out pixel matching on each frame of a video by utilizing an improved Gaussian mixture algorithm to obtain a corresponding static background and a corresponding dynamic foreground, and then extracts the dynamic foreground according to a certain rule to obtain a coarsely screened smoke foreground; the second part mainly adopts a YOLOv3 target detection algorithm to extract smoke characteristics and carry out secondary fine screening; when the smoke features are extracted by using YOLOv3, the preprocessed smoke data set is used as the input of a model, repeated training and parameter adjustment are carried out in the model, finally, the weight for representing the smoke features is generated, calling is carried out in the smoke foreground obtained by rough screening, secondary fine screening is carried out on the smoke region, and the specific position and range of the video smoke are determined finally, so that the smoke is effectively detected;
(1.1) improving mixed Gaussian smoke foreground extraction:
(1.1.1) matching pixel points of a current video frame with Gaussian distribution, if the formula (1-2) is satisfied, matching the pixel points with the Gaussian distribution, and if the formula is not satisfied, indicating that the pixel points are not matched;
|Xi,t-μi|＜2.5σi (1-2)
(1.1.2) if the matching is successful, updating the Gaussian background model by using the pixel point, wherein the updated formula is represented by a formula (1-3), a formula (1-4) and a formula (1-5):
ωi,t＝(1-α)ωi-1+α(Mi,t) (1-3)
μi,t＝(1-θ)μi,t-1+θXt (1-4)
σ2 i,t＝(1-θ)σ2 i,t-1+θ(Xt-μi,t-1)T(Xt-μi,t-1) (1-5)
wherein α is a learning rate; mi,tThe weight value change used for controlling a plurality of Gaussian distributions; when the matching is successful, Mi,t1 is ═ 1; when the matching fails, Mi,t0; alpha and theta may reflect the ability of gaussians to adapt to different scenarios;
(1.1.3) if the current distribution is not matched with any Gaussian distribution, adding a Gaussian distribution or updating the Gaussian distribution to replace the distribution with the minimum priority;
the K Gaussian distributions are arranged from high to low according to the priority, and are represented by the following formula (1-6):
(1.1.4) after updating parameters of the Gaussian model, arranging K Gaussian distributions of the pixel points according to the priority, and taking the first B Gaussian distributions as ideal description of background pixels:
in the formula, ωkRepresenting the weight of the Kth Gaussian function in one pixel; t is typically taken to be between 0.5 and 1, where an empirical value of 0.85 is taken, and continuing on for Xi,tPerforming matching test with the B Gaussian distributions, wherein the matching with any one of the previous B Gaussian distributions is the foreground, and otherwise, the matching is the background;
(1.2), YOLOv3 model and training:
predicting a target area and a target category in a neural network model; YOLOv3 adopts a new Darknet-53 network structure of ResNet idea, adds a residual error module in the network, and adopts 3 feature maps with different scales to detect objects, so that more fine-grained features can be detected;
step two, a smoke detection algorithm based on the improved YOLOv 3:
the channel attention mechanism is nested in the last layer of a Residual convolution Block Residul Block in a YOLOv3 network structure, in smoke target detection, when two or more target frames are close, the frame with lower score is deleted because of overlarge overlapping area, the threshold value of NMS needs to be manually determined, detection is missed when the threshold value is small, false detection is carried out when the threshold value is large, and non-maximum value inhibition is changed into Soft-NMS.
2. The deep learning based smoke detection method according to claim 1, wherein: in the smoke detection, the Soft-NMS firstly generates some smoke candidate areas, secondly obtains category scores through a classification network, meanwhile obtains more accurate position parameters through a regression network, and finally obtains the final detection result through non-maximum inhibition.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110542645.8A CN113537226A (en) | 2021-05-18 | 2021-05-18 | Smoke detection method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110542645.8A CN113537226A (en) | 2021-05-18 | 2021-05-18 | Smoke detection method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113537226A true CN113537226A (en) | 2021-10-22 |
Family
ID=78094678
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110542645.8A Pending CN113537226A (en) | 2021-05-18 | 2021-05-18 | Smoke detection method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113537226A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114132745A (en) * | 2021-11-30 | 2022-03-04 | 北京新风航天装备有限公司 | Automatic workpiece loading and unloading system and method based on AGV and machine vision |
| CN114550104A (en) * | 2022-04-22 | 2022-05-27 | 科大天工智能装备技术（天津）有限公司 | Fire monitoring method and system |
| CN116977634A (en) * | 2023-07-17 | 2023-10-31 | 应急管理部沈阳消防研究所 | Fire smoke detection method based on laser radar point cloud background subtraction |
| WO2024051297A1 (en) * | 2022-09-09 | 2024-03-14 | 南京邮电大学 | Lightweight fire smoke detection method, terminal device and storage medium |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105469105A (en) * | 2015-11-13 | 2016-04-06 | 燕山大学 | Cigarette smoke detection method based on video monitoring |
| WO2019048604A1 (en) * | 2017-09-09 | 2019-03-14 | Fcm Dienstleistungs Ag | Automatic early detection of smoke, soot and fire with increased detection reliability using machine learning |
| CN111079584A (en) * | 2019-12-03 | 2020-04-28 | 东华大学 | Rapid vehicle detection method based on improved YOLOv3 |
| CN111553387A (en) * | 2020-04-03 | 2020-08-18 | 上海物联网有限公司 | Yolov 3-based personnel target detection method |
| WO2020206861A1 (en) * | 2019-04-08 | 2020-10-15 | 江西理工大学 | Yolo v3-based detection method for key object at transportation junction |
| CN112070713A (en) * | 2020-07-03 | 2020-12-11 | 中山大学 | Multi-scale target detection method introducing attention mechanism |
| CN112101434A (en) * | 2020-09-04 | 2020-12-18 | 河南大学 | Infrared image weak and small target detection method based on improved YOLO v3 |
| CN112434672A (en) * | 2020-12-18 | 2021-03-02 | 天津大学 | Offshore human body target detection method based on improved YOLOv3 |
| CN112634283A (en) * | 2020-12-22 | 2021-04-09 | 华中科技大学 | Hip joint segmentation model establishment method using small sample image training and application thereof |
-
2021
- 2021-05-18 CN CN202110542645.8A patent/CN113537226A/en active Pending
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105469105A (en) * | 2015-11-13 | 2016-04-06 | 燕山大学 | Cigarette smoke detection method based on video monitoring |
| WO2019048604A1 (en) * | 2017-09-09 | 2019-03-14 | Fcm Dienstleistungs Ag | Automatic early detection of smoke, soot and fire with increased detection reliability using machine learning |
| WO2020206861A1 (en) * | 2019-04-08 | 2020-10-15 | 江西理工大学 | Yolo v3-based detection method for key object at transportation junction |
| CN111079584A (en) * | 2019-12-03 | 2020-04-28 | 东华大学 | Rapid vehicle detection method based on improved YOLOv3 |
| CN111553387A (en) * | 2020-04-03 | 2020-08-18 | 上海物联网有限公司 | Yolov 3-based personnel target detection method |
| CN112070713A (en) * | 2020-07-03 | 2020-12-11 | 中山大学 | Multi-scale target detection method introducing attention mechanism |
| CN112101434A (en) * | 2020-09-04 | 2020-12-18 | 河南大学 | Infrared image weak and small target detection method based on improved YOLO v3 |
| CN112434672A (en) * | 2020-12-18 | 2021-03-02 | 天津大学 | Offshore human body target detection method based on improved YOLOv3 |
| CN112634283A (en) * | 2020-12-22 | 2021-04-09 | 华中科技大学 | Hip joint segmentation model establishment method using small sample image training and application thereof |
Non-Patent Citations (3)
| Title |
|---|
| 张俊鹏: "基于深度学习的工业烟尘排放目标分割方法研究", 《中国优秀硕士学位论文全文数据库 工程科技I辑》 * |
| 程淑红 等: "改进的混合高斯与YOLOv2融合烟雾检测算法", 《计量学报》 * |
| 马继勇: "改进的混合高斯与YOLOv2融合烟雾检测算法", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114132745A (en) * | 2021-11-30 | 2022-03-04 | 北京新风航天装备有限公司 | Automatic workpiece loading and unloading system and method based on AGV and machine vision |
| CN114550104A (en) * | 2022-04-22 | 2022-05-27 | 科大天工智能装备技术（天津）有限公司 | Fire monitoring method and system |
| CN114550104B (en) * | 2022-04-22 | 2022-08-05 | 科大天工智能装备技术（天津）有限公司 | Fire monitoring method and system |
| WO2024051297A1 (en) * | 2022-09-09 | 2024-03-14 | 南京邮电大学 | Lightweight fire smoke detection method, terminal device and storage medium |
| CN116977634A (en) * | 2023-07-17 | 2023-10-31 | 应急管理部沈阳消防研究所 | Fire smoke detection method based on laser radar point cloud background subtraction |
| CN116977634B (en) * | 2023-07-17 | 2024-01-23 | 应急管理部沈阳消防研究所 | Fire smoke detection method based on laser radar point cloud background subtraction |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113537226A (en) | Smoke detection method based on deep learning | |
| CN107609470B (en) | Method for detecting early smoke of field fire by video | |
| CN110688925B (en) | Cascade target identification method and system based on deep learning | |
| Chaturvedi et al. | A survey on vision-based outdoor smoke detection techniques for environmental safety | |
| CN114863097B (en) | Infrared dim target detection method based on attention mechanism convolutional neural network | |
| CN101515326A (en) | Method for identifying and detecting fire flame in big space | |
| CN109919073B (en) | Pedestrian re-identification method with illumination robustness | |
| CN112070070B (en) | LW-CNN method and system for urban remote sensing scene recognition | |
| WO2024051297A1 (en) | Lightweight fire smoke detection method, terminal device and storage medium | |
| US20240005759A1 (en) | Lightweight fire smoke detection method, terminal device, and storage medium | |
| CN115240075B (en) | Construction and training method of electric power vision multi-granularity pre-training large model | |
| CN111582074A (en) | Monitoring video leaf occlusion detection method based on scene depth information perception | |
| CN114359733A (en) | Vision-based smoke fire detection method and system | |
| CN112862150A (en) | Forest fire early warning method based on image and video multi-model | |
| CN114580541A (en) | Fire disaster video smoke identification method based on time-space domain double channels | |
| CN116071685A (en) | Thermal power plant smoke and fire detection method based on inter-frame difference method and improved YOLOv4 | |
| CN114399734A (en) | Forest fire early warning method based on visual information | |
| CN117523437A (en) | Real-time risk identification method for substation near-electricity operation site | |
| CN112488213A (en) | Fire picture classification method based on multi-scale feature learning network | |
| CN110334703B (en) | Ship detection and identification method in day and night image | |
| CN109409224B (en) | Method for detecting flame in natural scene | |
| CN110751014A (en) | Flame detection system and method | |
| CN114998711A (en) | Method and system for detecting aerial infrared small and weak target and computer storage medium | |
| CN114998801A (en) | Forest fire smoke video detection method based on contrast self-supervision learning network | |
| CN114332754A (en) | Cascade R-CNN pedestrian detection method based on multi-metric detector |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20211022 |
|
| WD01 | Invention patent application deemed withdrawn after publication |