A method of smoke detection and a smoke detection system

Abstract

The application discloses a smoke detection method, comprising the following steps: installation and setting of a smoke detection system; smoke area extraction; processing of a smoke area picture; processing of a smoke mask picture with a transparent background; calling, synthesizing and labeling; combining and overall planning to form a data set; modeling training; smoke detection. The application also discloses a smoke detection system. The application analyzes and calculates motion trajectory features, area size, color and other related vector statistical features of smoke, obtains motion feature data, feeds back the motion feature data to a processing module for processing, guarantees the accuracy and precision of smoke area extraction, improves the effect and efficiency of subsequent processing, and guarantees the detection precision. The application sets a calling detection module to compare, predict and detect, realizes smoke detection on one picture in a video, does not depend on real-time video signals, saves a large amount of bandwidth and flow resources, and greatly reduces the cost.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smoke detection prediction, and in particular to a smoke detection method and a smoke detection system. BACKGROUND

[0002] Fire is one of the disasters that threaten human life and property safety at present, from small to large, such as personal home, building fire, and forest fire. For example, fire caused by improper use of electricity and fire, fire caused by failure of household electrical appliances, and natural fire caused by lightning in forest. Therefore, it is very important to discover fire in time and give an alarm. At present, the light sensor, smoke sensor, temperature sensor and other devices can be used to realize the early warning of fire.

[0003] With the continuous development of computer vision technology, smoke recognition algorithm is also used to analyze the image collected by the camera to determine whether there is smoke, and further determine whether there is fire, so that fire warning can be given in the early stage of fire.

[0004] However, the commonly used smoke recognition algorithm needs to extract features from the image analysis and processing in real time to determine whether there is smoke, that is, to determine and predict through real-time video. The transmission of video signal needs to consume a large amount of bandwidth and flow resources, especially in the monitoring scene where smoke does not often occur. The consumption of bandwidth and flow is huge for long-time uninterrupted video signal transmission, and the cost is high. Moreover, the existing smoke detection method uses fDSST tracking algorithm to automatically label the smoke target. This method can automatically label the video frame picture, which can greatly reduce the labeling time and effort. However, since the picture content between adjacent video frames changes very little, a large number of similar training pictures will be collected. These similar pictures are not helpful to the improvement of the model effect, and are prone to overfitting phenomenon, which affects the detection accuracy and effect. Moreover, the existing smoke detection uses a trained model for judgment and prediction detection. However, the training sample of the existing smoke detection method has poor randomness, and the richness and difference of the data set are low, which affects the synthesis effect and recognition effect in actual scene, and affects the detection efficiency and accuracy. The existing smoke detection system has no classification function, which affects the response speed and accuracy of subsequent retrieval comparison, and affects the judgment efficiency, thereby affecting the recognition efficiency and time. Moreover, the existing smoke detection system and method have poor extraction accuracy of smoke area, which affects the smoke analysis and processing effect, and affects the synthesis effect and efficiency of the subsequent smoke training set, and affects the detection efficiency. SUMMARY

[0005] The present application is aimed at the technical deficiencies of the prior art, and provides a smoke detection method.

[0006] The application also discloses a smoke detection system.

[0007] The application adopts the technical scheme that is:

[0008] A smoke detection method comprises the following steps:

[0009] (1) Installation and setting of the smoke detection system: the smoke detection system is installed on a terminal, the smoke detection system is linked with a video system, the smoke detection system comprises a processing module, an intercepting and extracting module, a separation processing module, a processing and impurity removing module, an image adjustment processing module, a calling and synthesizing module, a synthesizing and labeling module, a combination and planning module, a saving module, a modeling module and a calling and detecting module, the saving module comprises a smoke mask picture saving module, a training set saving module, a background picture saving module, a synthesized picture saving module and a model saving module, and the calling and detecting module is provided with an environment recognizing module, a quick calling module and a comparison module;

[0010] (2) Smoke area extraction: the processing module controls the intercepting and extracting module to extract a smoke area in a video segment in the video system, and a smoke area picture is obtained;

[0011] (3) Processing of the smoke area picture: the processing and impurity removing module removes the interference of impurity parameters outside the smoke by performing image processing operations such as filtering, expansion and corrosion on the smoke area picture, then intercepts and saves the smoke part of the processed smoke area picture as a background transparent smoke mask picture, and the background transparent smoke mask picture is transmitted to the smoke mask picture saving module for archiving and saving;

[0012] (4) Processing of the background transparent smoke mask picture: the image adjustment processing module is provided with a calling module, the calling module calls the background transparent smoke mask picture in the smoke mask picture saving module, calls the background transparent smoke mask picture to the image adjustment processing module for processing, the image adjustment processing module performs random data enhancement operations such as flipping, rotating, scaling, color jittering, brightness adjusting and saturation adjusting on the background transparent smoke mask picture, and an enhanced mask picture is obtained;

[0013] (5) Calling, synthesizing and labeling: the calling and synthesizing module randomly calls a background picture, synthesizes the enhanced mask picture and the called background picture, obtains a synthesized picture, labels the synthesized picture through the synthesizing and labeling module, and after the labeling is completed, a synthesized labeled picture and a corresponding label file are formed, the synthesized labeled picture and the corresponding label file are saved to the synthesized picture saving module, and the label file is used to save the rectangular coordinate position of the smoke area in the synthesized labeled picture;

[0014] (6)Combination of the data set: the combination of the module for calling the synthesis of the picture saving module in the synthesis of the label picture, call out a specific number of different background synthesis of the label picture and the corresponding label file, then the multiple different synthesis of the label picture and the corresponding label file are combined and trained to form a synthesis of the label data set, and the different synthesis of the label data set is saved to the training set saving module;

[0015] (7) Modeling training: the modeling module trains the smoke detection model based on the deep learning target detection algorithm, trains and merges multiple labeled data sets in the training set saving module to form different smoke detection models and the corresponding environmental feature parameters of the smoke detection model, and the smoke detection model and the corresponding environmental feature parameters of the smoke detection model are archived to the model saving module.

[0016] (8) Smoke detection: the detection module quickly identifies the environment of a single picture in the video system to obtain the environmental feature parameters of the single picture, calculates the similarity between the environmental feature parameters of the single picture and the environmental feature parameters of the smoke detection model in the model saving module, and then the smoke detection model with high similarity is quickly retrieved from the model saving module through the quick retrieval module. The comparison module retrieves the smoke detection model to predict and detect the single picture in the video system, feeds back the detection and recognition results after the detection is completed, realizes the detection of the smoke in the picture, and completes the whole detection process.

[0017] Further improvement, the step (2) further comprises the following steps:

[0018] (2.1) The interception and extraction module intercepts and extracts multiple video segments in the video system, the interception and extraction module is provided with a background modeling module, the background modeling module performs background modeling processing on the first few frames of pictures in the video segment based on the background modeling technology, and the video background picture is obtained after the background modeling is completed.

[0019] (2.2) The interception and extraction module is also provided with an analysis module, the analysis module uses the motion characteristic interception technology to analyze each frame of picture containing smoke and the video background picture obtained by background modeling;

[0020] (2.3) The motion characteristic interception technology uses difference analysis method to analyze and calculate the motion trajectory characteristics, area size, color and other related vector statistical characteristics of smoke respectively, obtains motion characteristic data, and the interception and extraction module processes the motion characteristic data to obtain the smoke area picture.

[0021] Further improvement, the background modeling technology is one of the mixed Gaussian modeling method, VIBE modeling method and other background modeling methods.

[0022] As a further improvement, the step (5) further comprises the following steps: the calling and synthesizing module is provided with a background picture random calling module, the calling and synthesizing module calls different background pictures and synthesizes with the enhanced mask picture to obtain a plurality of synthesized marking pictures with different backgrounds and corresponding marking files.

[0023] As a further improvement, the step (8) further comprises the following steps:

[0024] (8.1) Environment parameter identification; the environment identification module identifies a single picture in the video system, the environment identification module separates the environment feature parameters and the environment feature quantities in the single picture, and generates a feature vector of the environment feature parameters and a feature vector of the environment feature quantities according to the separated environment feature parameters and the environment feature quantities,

[0025] (8.2) Smoke detection model calling of corresponding environment feature parameters: the fast calling module calls the smoke detection model of the feature vector of the consistent environment feature parameters and the feature vector of the environment feature quantities, and transmits the related data of the called smoke detection model to the comparison module for similarity judgment;

[0026] (8.3) Comparison of the comparison module: the comparison module is provided with a feature vector calculation module and an identification and judgment module, the feature vector calculation module compares the feature vector of the environment feature parameters and the feature vector of the environment feature quantities in the single picture with the feature vector of the environment feature parameters and the feature vector of the environment feature quantities in the smoke detection model, and obtains the difference between the two feature vectors.

[0027] (8.4) Identification and judgment: a set value is set in the identification and judgment module to judge the threshold of similarity, the identification module judges whether the difference between the two feature vectors is greater than the threshold according to the difference between the two feature vectors, and obtains the similarity information between the two according to the judgment result. Specifically, if it is greater than the threshold, it is determined that the two are not similar, otherwise, the two are similar.

[0028] As a further improvement, the feature quantity includes smoke area range, smoke area color difference and smoke area orientation, wherein the smoke area range includes height difference, width difference and area, the smoke area color difference includes weak color difference area, moderate color difference area and deep color difference area, and the smoke area orientation includes smoke area orientation trend and environment reference orientation trend.

[0029] As a further improvement, the deep learning target detection algorithm in the step (7) is one of Faster RCNN detection algorithm, SSD detection algorithm, YOLO detection algorithm and other target detection algorithms.

[0030] A kind of smoke detection system of smoke detection method, the training set saving module is provided with classification saving module, the classification saving module includes sunny day environment saving module, rainy day environment saving module, foggy environment saving module and dim environment saving module.

[0031] Further improvement, the processing module is linked with interception extraction module, separation processing module, processing impurity module, image adjustment processing module, call synthesis module, synthesis labeling module, combination planning module and saving module, the processing module is also provided with import module and labeled data set call module, the import module is used for the import of background picture, the labeled data set call module is used for the quick call action of synthesis labeled data set.

[0032] Further improvement, the image adjustment processing module is provided with random synthesis module, and the random synthesis module is used for random parameter setting action to the smoke mask picture of background transparency.

[0033] The beneficial effects of the present application are: the present application uses the motion characteristic interception technology constituted by difference analysis method to analyze and calculate the motion trajectory characteristics, region size, color and other related vector statistical characteristics of smoke, obtain motion characteristic data, and feed back the motion characteristic data to the processing module for processing, to ensure the accuracy and precision of smoke region extraction, improve the effect and efficiency of subsequent processing, and ensure the detection accuracy of subsequent processing; the processing impurity module is set to filter, dilate, erode and other image processing operations on the smoke region picture, to remove the interference of impurity parameters outside the smoke, and the image adjustment processing module is set to perform random data enhancement operation on the smoke mask picture of background transparency, so that the smoke mask picture has great richness and difference, and is not easy to overfit, thereby ensuring the data amount of subsequent training set, improving the detection accuracy and effect of subsequent processing; the background picture random call module and the background picture saving module are set to improve the randomness and data amount of the synthesized background, improve the richness and difference of the data set, and make the synthesized picture more suitable for actual application scenarios, improve the synthesis effect in subsequent processing and recognition effect in actual scenarios, and improve the detection efficiency and accuracy in subsequent processing; the combination planning module is set to call the synthesis labeled picture in the synthesis picture saving module, call a specific number of synthesis labeled pictures with different backgrounds, then combine and train multiple synthesis labeled pictures, constitute a synthesis labeled data set, and save it to the training set saving module, thereby improving the data amount of the training set, and providing a strong data basis for the training of the smoke detection model.

[0034] The application avoids extremely tedious labeling work, greatly improves the algorithm development efficiency, effectively guarantees the detection effect of the model, guarantees the detection precision, realizes classification saving through the setting of the classification saving module, the classification saving module includes a sunny environment saving module, a rainy environment saving module, a haze environment saving module and a dim environment saving module, the classification saving is convenient for the rapid retrieval of the subsequent training set, improves the response speed and accuracy of the retrieval comparison, improves the judgment efficiency, improves the efficiency and time of identification.

[0035] The application is further described below in combination with the drawings and specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0036] Fig. 1 It is a flowchart of the smoke detection method of the embodiment;

[0037] Fig. 2 It is a module schematic diagram of the smoke detection system of the embodiment. DETAILED DESCRIPTION

[0038] The following description is only a preferred embodiment of the application, and does not limit the protection scope of the application.

[0039] Embodiment, see attached Figs. 1-2 A smoke detection method, comprising the following steps:

[0040] (1) Installation and setting of the smoke detection system: install the smoke detection system on the terminal, the smoke detection system is linked with the video system, the smoke detection system includes a processing module, an interception extraction module, a separation processing module, a processing impurity removal module, an image adjustment processing module, a retrieval synthesis module, a synthesis labeling module, a combination planning module, a saving module, a modeling module and a retrieval detection module, the saving module includes a smoke mask picture saving module, a training set saving module, a background picture saving module, a synthesis picture saving module and a model saving module, the retrieval detection module is provided with an environment recognition module, a rapid retrieval module and a comparison module;

[0041] (2) Smoke area extraction: the processing module controls the interception extraction module to extract the smoke area in the video segment in the video system, and obtains a smoke area picture;

[0042] (3) Processing of the smoke area picture: the processing impurity removal module removes the interference of impurity parameters outside the smoke area by performing image processing operations such as filtering, expansion and corrosion on the smoke area picture, then intercepts and saves the smoke part of the processed smoke area picture as a background transparent smoke mask picture, and transmits the background transparent smoke mask picture to the smoke mask picture saving module for archiving and saving;

[0043] (4) Processing of background transparent smoke mask pictures: the image adjustment processing module is provided with a calling module, which calls the background transparent smoke mask pictures in the smoke mask picture saving module, calls to the image adjustment processing module for processing, and the image adjustment processing module performs random data enhancement operations such as flipping, rotating, scaling, color jittering, brightness adjustment, saturation adjustment, etc. on the background transparent smoke mask pictures to obtain enhanced mask pictures;

[0044] (5) Calling, synthesizing and labeling: the calling and synthesizing module randomly calls background pictures, synthesizes the enhanced mask pictures with the called background pictures to obtain synthesized pictures, and labels the synthesized pictures through the synthesis labeling module. The labeled pictures and the corresponding labeling files are saved to the synthesized picture saving module. The labeling files are used to save the rectangular coordinate positions of the smoke areas in the synthesized pictures.

[0045] (6) Combining and planning to constitute a data set: the combining and planning module calls the synthesized pictures in the synthesized picture saving module, calls out a specific number of synthesized pictures with different backgrounds and the corresponding labeling files, and then combines and plans multiple synthesized pictures and the corresponding labeling files to constitute a synthesized labeling data set, and saves the constituted different synthesized labeling data sets to the training set saving module.

[0046] (7) Modeling and training: the modeling module trains a smoke detection model based on a deep learning target detection algorithm, trains and combines multiple labeling data sets in the training set saving module to constitute different smoke detection models and the corresponding environmental feature parameters of the smoke detection models, and archives the constituted smoke detection models and the corresponding environmental feature parameters of the smoke detection models to the model saving module.

[0047] (8) Smoke detection: the calling and detecting module quickly identifies the environment of a single picture in a video system to obtain the environmental feature parameters of the single picture, calculates the similarity between the environmental feature parameters of the single picture and the environmental feature parameters of the smoke detection model in the model saving module, calls the smoke detection model with high similarity from the model saving module through the fast calling module after the calculation, and the comparison module calls the smoke detection model to predict and detect the single picture in the video system after the calling, feeds back the detection and recognition results after the detection, realizes the detection of the smoke in the picture, and completes the whole detection process.

[0048] The step (2) further comprises the following steps:

[0049] (2.1) the intercepting and extracting module performs intercepting and extracting actions on a plurality of video segments in a video system, the intercepting and extracting module is provided with a background modeling module, the background modeling module performs background modeling processing on the first few frames of pictures in the video segment based on a background modeling technology, and a video background picture is obtained after the background modeling is completed;

[0050] (2.2) the intercepting and extracting module is also provided with an analysis module, the analysis module uses a motion characteristic intercepting technology to analyze each frame of picture containing smoke and the video background picture obtained by background modeling;

[0051] (2.3) the motion characteristic intercepting technology uses a difference analysis method to analyze and calculate the motion trajectory characteristics, region size, color and other related vector statistical characteristics of the smoke, to obtain motion characteristic data, and the intercepting and extracting module processes the motion characteristic data to obtain a smoke region picture.

[0052] Further improvement, the background modeling technology is one of a mixed Gaussian modeling method, a VIBE modeling method and other background modeling methods.

[0053] The step (5) further comprises the following steps, the calling and synthesizing module is provided with a background picture random calling module, the calling and synthesizing module calls different background pictures and synthesizes with the enhanced mask picture to obtain a plurality of synthesized marking pictures with different backgrounds and corresponding marking files.

[0054] The step (8) further comprises the following steps:

[0055] (8.1) environmental parameter identification; the environment identification module identifies a single picture in the video system, the environment identification module separates the environmental characteristic parameters and the environmental characteristic quantities in the single picture, generates a feature vector of the environmental characteristic parameters and a feature vector of the environmental characteristic quantities according to the separated environmental characteristic parameters and the environmental characteristic quantities,

[0056] (8.2) smoke detection model calling of corresponding environmental characteristic parameters: the fast calling module calls the smoke detection model of the feature vector of the consistent environmental characteristic parameters and the feature vector of the environmental characteristic quantities, and transmits the related data of the called smoke detection model to the comparison module for similarity judgment;

[0057] (8.3) comparison of the comparison module: the comparison module is provided with a feature vector calculation module and an identification and judgment module, the feature vector calculation module compares the feature vector of the environmental characteristic parameters and the feature vector of the environmental characteristic quantities in the single picture with the feature vector of the environmental characteristic parameters and the feature vector of the environmental characteristic quantities in the smoke detection model, and obtains the difference of the two feature vectors.

[0058] (8.4) Recognition judgment: a set value is set in the recognition judgment module to judge the threshold of similarity, and the recognition module judges according to the difference between the two feature vectors. Whether the difference between the two feature vectors is greater than the threshold value, according to the judgment result, the similarity information between the two is obtained. Specifically, if it is greater than the threshold value, it is determined that the two are not similar, otherwise, the two are similar.

[0059] The feature quantity includes smoke area range, smoke area color difference and smoke area orientation, wherein the smoke area range includes height difference, width difference and area, the smoke area color difference includes weak color difference area, moderate color difference area and deep color difference area, and the smoke area orientation includes smoke area orientation trend and environment reference orientation trend.

[0060] The step (7) is based on a deep learning target detection algorithm, which is one of Faster RCNN detection algorithm, SSD detection algorithm, YOLO detection algorithm and other target detection algorithms.

[0061] A smoke detection system for implementing a smoke detection method, wherein the training set saving module is provided with a classification saving module, and the classification saving module includes a sunny day environment saving module, a rainy day environment saving module, a haze environment saving module and a dim environment saving module.

[0062] The processing module is linked and communicated between the interception and extraction module, the separation processing module, the processing and impurity removal module, the image adjustment processing module, the calling and synthesis module, the synthesis labeling module, the combination and planning module and the saving module. The processing module is also provided with an import module and a labeled data set calling module. The import module is used for importing background pictures, and the labeled data set calling module is used for quick calling action of synthesized labeled data set.

[0063] The image adjustment processing module is provided with a random synthesis module, which is used for random parameter setting action of background transparent smoke mask picture.

[0064] The application adopts the motion characteristic interception technology formed by the differential analysis method, analyzes and calculates the motion trajectory characteristics, regional size, color and other related vector statistical characteristics of the smoke respectively, obtains the motion characteristic data, feeds back the motion characteristic data to the processing module for processing, ensures the accuracy and precision of the smoke area extraction, improves the effect and efficiency of the subsequent processing, and ensures the detection accuracy of the subsequent processing; the processing impurity module is set to filter, dilate, erode and other image processing operations on the smoke area picture, remove the interference of impurity parameters outside the smoke, and the image adjustment processing module is set to perform random data enhancement operation on the smoke mask picture with transparent background, so that the smoke mask picture has great richness and difference, and overfitting is not easy to occur, thereby ensuring the data amount of the subsequent training set, improving the detection accuracy and effect of the subsequent processing; the background picture random calling module and the background picture saving module are set to improve the randomness and data amount of the synthesized background, improve the richness and difference of the data set, and make the synthesized picture more suitable for the actual application scene, improve the synthesis effect and recognition effect in the actual scene, and improve the detection efficiency and accuracy of the subsequent processing; the combination planning module is set to call the synthesized labeled pictures in the synthesized picture saving module, call a specific number of synthesized labeled pictures with different backgrounds, then combine and plan the plurality of different synthesized labeled pictures, form a synthesized labeled data set, and save to the training set saving module, thereby improving the data amount of the training set, and providing a strong data basis for the training of the smoke detection model.

[0065] The application avoids the extremely tedious labeling work, greatly improves the algorithm development efficiency, effectively guarantees the detection effect of the model, and guarantees the detection accuracy; the classification saving module is set to realize classification saving, the classification saving module includes a sunny environment saving module, a rainy environment saving module, a haze environment saving module and a dim environment saving module, the classification saving facilitates the rapid calling of the subsequent training set, improves the response speed and accuracy of the calling comparison, improves the judgment efficiency, and improves the efficiency and time of identification.

[0066] The application sets the calling detection module to perform comparison prediction and detection, the calling detection module is provided with an environment recognition module, a rapid calling module and a comparison module, a smoke model is established to detect the smoke in a picture of the video, and the smoke detection does not depend on the real-time video signal, a large amount of bandwidth and flow resources are saved, the cost is greatly reduced, the smoke detection is used to realize rapid classification comparison, the accuracy and effect of detection and prediction are greatly improved, and the judgment efficiency and accuracy are improved.

[0067] The application is not limited to the above-mentioned embodiments, other smoke detection methods and smoke detection systems obtained by using the same or similar structures, devices, processes or methods as the above-mentioned embodiments of the application are also within the protection scope of the application.

Claims

1. A method of detecting smoke, characterized by: It comprises the following steps: (1) Installation and setting of smoke detection system: install the smoke detection system on the terminal, the smoke detection system is linked with the video system, the smoke detection system comprises a processing module, an intercepting and extracting module, a separation processing module, a processing and impurity removing module, an image adjustment processing module, a calling and synthesizing module, a synthesizing labeling module, a combination and planning module, a saving module, a modeling module and a calling and detecting module, the saving module comprises a smoke mask picture saving module, a training set saving module, a background picture saving module, a synthesized picture saving module and a model saving module, the calling and detecting module is provided with an environment recognition module, a quick calling module and a comparison module; (2) Smoke area extraction: the processing module controls the intercepting and extracting module to extract the smoke area in the video segment in the video system, and the smoke area picture is obtained; (3) Processing of smoke area picture: the processing and impurity removing module removes the impurities in the extracted smoke area picture, first carries out image processing operations such as filtering, expansion and corrosion on the smoke area picture, removes the interference of impurity parameters outside the smoke, then intercepts and saves the smoke part of the processed smoke area picture as a background transparent smoke mask picture, and transmits the background transparent smoke mask picture to the smoke mask picture saving module for archiving; (4) Processing of background transparent smoke mask picture: the image adjustment processing module is provided with a calling module, the calling module calls the background transparent smoke mask picture in the smoke mask picture saving module, calls to the image adjustment processing module for processing, the image adjustment processing module carries out random data enhancement operations such as flipping, rotating, scaling, color jitter, brightness adjustment, saturation adjustment, etc. on the background transparent smoke mask picture, and obtains an enhanced mask picture; (5) Calling, synthesizing and labeling: the calling and synthesizing module randomly calls a background picture, synthesizes the enhanced mask picture with the called background picture, obtains a synthesized picture, labels the synthesized picture through the synthesizing labeling module, and labels to form a synthesized labeling picture and a corresponding labeling file. Save the synthesized labeling picture and the corresponding labeling file to the synthesized picture saving module, and the labeling file is used to save the rectangular coordinate position of the smoke area in the synthesized labeling picture; (6) Combination and planning to form a data set: the combination and planning module calls the synthesized labeling pictures in the synthesized picture saving module, calls out a specific number of synthesized labeling pictures with different backgrounds and corresponding labeling files, then combines and plans a plurality of different synthesized labeling pictures and corresponding labeling files for training, forms a synthesized labeling data set, and saves the formed different synthesized labeling data sets to the training set saving module; (7) modeling training: the modeling module trains the smoke detection model based on a deep learning target detection algorithm, merges a plurality of labeled data sets in a training set storage module to form different smoke detection models and their corresponding environmental feature parameters of the smoke detection models, and archives the smoke detection models and their corresponding environmental feature parameters of the smoke detection models to a model storage module; (8) smoke detection: the detection module quickly identifies the environment of a single picture in a video system to obtain the environmental feature parameters of the single picture, calculates the similarity between the environmental feature parameters of the single picture and the environmental feature parameters of the smoke detection model in the model storage module, and after the calculation, the smoke detection model with high similarity is quickly retrieved from the model storage module through the quick retrieval module. The comparison module compares the smoke detection model with the single picture in the video system to predict and detect the single picture, feeds back the detection and recognition results after the detection is completed, detects the smoke in the picture, and completes the entire detection process.

2. The smoke detection method of claim 1, wherein: The step (2) further comprises the following steps: (2.1) The extraction module extracts a plurality of video segments in the video system, and the extraction module is provided with a background modeling module. The background modeling module performs background modeling processing on the first few frames of pictures in the video segment based on a background modeling technology, and obtains a video background picture after the background modeling is completed. (2.2) The extraction module is also provided with an analysis module. The analysis module uses a motion characteristic extraction technology to analyze each frame of picture containing smoke and the video background picture obtained by background modeling. (2.3) The motion characteristic extraction technology uses a difference analysis method to analyze and calculate the motion trajectory characteristics, region size, color and other related vector statistical characteristics of the smoke, to obtain motion characteristic data, and the extraction module processes the motion characteristic data to obtain a smoke region picture.

3. The method of smoke detection according to claim 2, wherein: The background modeling technology is one of a mixed Gaussian modeling method, a VIBE modeling method and other background modeling methods.

4. The smoke detection method of claim 1, wherein: The step (5) further comprises the following steps. The retrieval and synthesis module is provided with a background picture random retrieval module. The retrieval and synthesis module retrieves different background pictures and synthesizes them with the enhanced mask picture to obtain a plurality of synthetic labeled pictures with different backgrounds and their corresponding labeled files.

5. The smoke detection method of claim 1, wherein: The step (8) further comprises the following steps: (8.1) Environmental parameter identification: the environmental identification module identifies a single picture in the video system. The environmental identification module separates the environmental feature parameters and the environmental feature quantities in the single picture, generates a feature vector of the environmental feature parameters and a feature vector of the environmental feature quantities according to the separated environmental feature parameters and environmental feature quantities, (8.2) Smoke detection model retrieval of corresponding environmental feature parameters: the quick retrieval module retrieves the smoke detection model of the feature vector of the consistent environmental feature parameters and the feature vector of the environmental feature quantities, and transmits the related data of the retrieved smoke detection model to the comparison module for similarity judgment. (8.3) Comparison module comparison: the comparison module is provided with a feature vector calculation module and an identification and judgment module, the feature vector calculation module compares the feature vectors of the environmental feature parameters and the feature vectors of the environmental feature quantities in a single picture with the feature vectors of the environmental feature parameters and the feature vectors of the environmental feature quantities in the smoke detection model, and obtains the modulus of the difference between the two feature vectors; (8.4) Identification and judgment: a set value is set in the identification and judgment module to judge the threshold of similarity, the identification module judges whether the modulus of the difference between the two feature vectors is greater than the threshold according to the modulus of the difference between the two feature vectors, and obtains the similarity information between the two according to the judgment result, specifically, if it is greater than the threshold, it is determined that the two are not similar, otherwise, the two are similar.

6. The method of smoke detection according to claim 5, wherein: The feature quantity includes smoke area range, smoke area color difference and smoke area orientation, wherein the smoke area range includes height difference, width difference and area, the smoke area color difference includes weak color difference area, moderate color difference area and deep color difference area, and the smoke area orientation includes smoke area orientation trend and environment reference orientation trend.

7. The method of claim 1, wherein: The step (7) is based on a deep learning target detection algorithm, which is one of Faster RCNN detection algorithm, SSD detection algorithm, YOLO detection algorithm and other target detection algorithms.

8. A smoke detection system implementing the method of smoke detection according to any one of claims 1 to 7, characterized in that: The training set saving module is provided with a classification saving module, which includes a sunny environment saving module, a rainy environment saving module, a haze environment saving module and a dim environment saving module.

9. The smoke detection system of claim 8, wherein: The processing module is linked and communicated between the interception and extraction module, the separation processing module, the processing and impurity removal module, the image adjustment processing module, the calling and synthesis module, the synthesis labeling module, the combination and planning module and the saving module, the processing module is also provided with an import module and a labeled data set calling module, the import module is used for importing background picture, and the labeled data set calling module is used for quick calling action of synthesis labeled data set.

10. The smoke detection system of claim 9, wherein: The image adjustment processing module is provided with a random synthesis module, which is used for random parameter setting action of background transparent smoke mask picture.

Images