Lightweight smoke detection method based on self-evolution of empirical model for forest fires

Abstract

The present invention discloses a lightweight smoke detection method for forest fires based on the self-evolution of an empirical model, which belongs to the field of target detection. The specific steps include: reading the image of the input video frame by frame, performing color space conversion on the original image, extracting a preliminary identification area in the image using an empirical model based on smoke color, and comparing it with the final identification area extracted at the previous moment, accurately identifying the changed image area using a deep learning algorithm, and realizing the self-evolution of the empirical model based on the recognition results.

Description

Technical Field

[0001] The present invention relates to the field of target detection, and in particular to a lightweight smoke detection method based on self-evolution of empirical models for forest fires. Background Art

[0002] The frequent occurrence of fire accidents seriously endangers public safety and causes huge losses. Therefore, fire detection has become an indispensable disaster prevention measure. Traditional fire detection equipment, such as traditional detectors with temperature and smoke sensing functions, is not only affected by the environment, but also has shortcomings such as limited detection range, high false alarm rate, and delayed response.

[0003] With the popularization of video surveillance, the development of deep learning and target detection technology, the technology of early warning fires through video surveillance has been widely studied and applied. However, due to the real-time characteristics of video, this monitoring and early warning technology often has a large lag and has relatively high requirements for hardware processing equipment.

[0004] The combination of mathematical model and deep learning is called digital-analog hybrid method. The model method is simple and not accurate enough in various scenarios, but it will not be outrageous. The data method (machine learning) is complex and very accurate in training scenarios, but it will be limited in some scenarios and will be very inaccurate. Based on the above characteristics, there are many ways to combine digital and analog. This patent is based on the following assumptions: the model method is simple but inaccurate; the data method is complex but accurate (for specific scenarios). Therefore, in order to achieve accurate and lightweight smoke detection, the two methods are combined, the model method is mainly used for detection, the data method is used for detection in uncertain areas, and the model method is corrected (self-evolution) according to the results, so as to ensure accuracy on the basis of overall lightweight, reduce the requirements for hardware processing equipment, and take into account timeliness and accuracy. Summary of the Invention

[0005] The purpose of the present invention is to provide a lightweight smoke detection method based on the self-evolution of empirical models for forest fires, which combines the correlation between frame images in the video, reduces the amount of data processed by hardware equipment, lowers hardware requirements, and solves the problem of efficient identification in fire monitoring systems.

[0006] To achieve the above functions, the present invention designs a lightweight smoke detection method based on the self-evolution of an empirical model for forest fires. The method performs the following steps S1 to S7 for a target forest area to detect whether smoke is generated in the target forest area and thus determine whether a fire has occurred:

[0007] Step S1. Real-time video of the target forest area is captured and converted into a time-sequential image sequence. Each frame of the image is divided into a preset number of image blocks, and then the process proceeds to step S2.

[0008] Step S2. For the first image frame in the image sequence obtained in step S1, a pre-trained deep learning model is used that takes each image block in the image as input and uses as output a final identification corresponding to each image block in the image, indicating whether each image block contains smoke. The model determines whether each image block in the first image frame contains smoke, where a final identification of 1 indicates that the corresponding image block contains smoke, and a final identification of 0 indicates that the corresponding image block does not contain smoke.

[0009] Then, starting with the second frame in the image sequence, the following steps S3 to S7 are performed for each frame in sequence to detect whether smoke is generated in the target forest area:

[0010] Step S3: Using the statistical histogram method of the empirical sample data, a conditional probability model is constructed based on the R, G, and B of the image, with the forest background as the occurrence event E0, the smoke foreground as the occurrence event E1, and other foregrounds as the occurrence event E2;

[0011] An empirical model is constructed based on the conditional probability model. According to the empirical model, each pixel in the image is divided into smoke foreground pixels, forest background pixels, and other foreground pixels. Each image block of the image is traversed and the number of smoke foreground pixels, forest background pixels, and other foreground pixels in each image block is counted.

[0012] Based on the number of smoke foreground pixels, forest background pixels, and other foreground pixels in the image block, determine whether the image block contains smoke and match the preliminary identification corresponding to the image block. A preliminary identification of 1 indicates that the corresponding image block contains smoke, and a preliminary identification of 0 indicates that the corresponding image block does not contain smoke.

[0013] Step S4. For each image block in the image, the preliminary identification is compared with the final identification of the image block at the same position in the adjacent previous image frame. If the preliminary identification of the subsequent image block in the two image blocks is different from the final identification of the previous image block, the subsequent image block is determined to be a pending confirmation area, and the process proceeds to step S5. Otherwise, the image block is determined to be a confirmed area, and the process proceeds to step S6 for the confirmed area.

[0014] Step S5. Input the image blocks of the area to be confirmed into the pre-trained deep learning model, output the final identification of the image blocks of the area to be confirmed, and based on the preliminary identification and final identification of the image blocks of the area to be confirmed, evolve and update the empirical model, traverse all image blocks of the area to be confirmed, complete the self-evolution of the empirical model, and then proceed to step S7;

[0015] Step S6. For each image block of the confirmed area, output the preliminary identification as the final identification, and then proceed to step S7;

[0016] Step S7: Based on the accumulation of judgment results of each image block, a pre-trained deep learning model is used to further verify each image block.

[0017] As a preferred technical solution of the present invention: the method of dividing each frame image into a preset number of image blocks in step S1 is as follows:

[0018] The size of each image in the image sequence is 256×256. For the image I at time t in the image sequence, t , divide it into image blocks of size 8×8, image I t The i-th image block in Contains 32×32 pixels, 1≤i≤64, Represents an image block The j-th pixel in , 1≤j≤1024, Respectively The red chroma color value, green chroma color value, and blue chroma color value.

[0019] As a preferred technical solution of the present invention, the specific steps of constructing the conditional probability model in step S3 are as follows:

[0020] Step S31. Construct conditional probability models based on the R, G, and B values ​​of the image, with the forest background as the occurrence event E0, the smoke foreground as the occurrence event E1, and other foregrounds as the occurrence event E2: P(E0|R), P(E1|R), P(E2|R), P(E0|G), P(E1|G), P(E2|G), P(E2|G), P(E1|B), P(E2|B), P(E2|B), where R represents the red chromaticity color value of the image, G represents the green chromaticity color value of the image, and B represents the blue chromaticity color value of the image;

[0021] Step S32: Construct an empirical model based on the conditional probability model, and classify each pixel that meets the following conditions as a smoke foreground pixel:

[0022]

[0023] Each pixel that meets the following conditions is classified as a forest background pixel:

[0024]

[0025] Where α1, α2, and α3 represent the importance coefficients corresponding to the conditional probabilities of R, G, and B, respectively, and Th1 is the joint conditional probability threshold. Represents image blocks Forest background countermeasure factor, smoke foreground countermeasure factor;

[0026] Each pixel that does not meet the above conditions is divided into other pixel points;

[0027] Step S33. Counting image blocks The number of forest background pixels Number of smoke foreground pixels Other pixel numbers If the following conditions are met:

[0028]

[0029] Then judge the image block For the smoke foreground area, the image block Matching preliminary identification Represents an image block There is smoke, otherwise the image block Matching preliminary identification Represents an image block There is no smoke in the scene; where Th2 is the pixel threshold of the smoke scene;

[0030] Step S34: Repeat step S33, traversing each image block until all image blocks in the image are judged.

[0031] As a preferred technical solution of the present invention: the method for determining the area to be confirmed in step S4 is as follows:

[0032]

[0033] Where, represents a modulo 2 addition operation, Represents the final identification of the i-th image block in the image at time t-1, if It means that the image block identification of the two frames before and after is different, that is, the image block is judged For the area to be confirmed, if It means that the image block identifiers of the two frames before and after are the same, that is, the image block is judged To confirm the area, For image blocks The final identification.

[0034] As a preferred technical solution of the present invention, the specific steps of step S5 are as follows:

[0035] Step S51. If the image block For areas to be confirmed, a pre-trained deep learning model is used to determine the image block Does it contain smoke? If the image block If there is smoke, If the image block is judged If there is no smoke,

[0036] Step S52: Image blocks of the area to be confirmed Based on preliminary identification Final logo The evolutionary update of the empirical model is as follows:

[0037]

[0038]

[0039] Where Δ is the evolution factor, Represents image blocks The forest background confrontation factor and the smoke foreground confrontation factor at the next moment;

[0040] Step S53: Repeat step S52 until all image blocks in the area to be confirmed are traversed and the self-evolution of the empirical model is completed.

[0041] As a preferred technical solution of the present invention: the specific method of further checking each image block using the pre-trained deep learning model in step S7 is as follows:

[0042] If the image I at time t in the image sequence t If there is no image block in the area to be confirmed, it is considered that the empirical model judgment has not changed, and the number of judgments of the empirical model CR increases by 1, otherwise CR increases by 0;

[0043] If any are all 0, 1≤i≤64, then when CR=N, where N is the preset number of judgments, and N is a positive integer, the number of smoke foreground pixels is selected For the image block with the most, a pre-trained deep learning model is used to determine whether the image block contains smoke, and the method described in step S5 is used to perform an evolutionary update of the empirical model;

[0044] If exists is 1, 1≤i≤64, then when CR=2N, select the number of smoke foreground pixels The least image block uses a pre-trained deep learning model to determine whether the image block contains smoke, and the method described in step S5 is used to evolve and update the empirical model.

[0045] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0046] The present invention designs a lightweight smoke detection method based on the self-evolution of an empirical model for forest fires. Smoke detection is performed using a model method, and a data method is used for detection in uncertain areas. The model method (self-evolution) is corrected based on the results, thereby ensuring accuracy while being lightweight overall. This method reduces the requirements for hardware processing equipment while taking into account both timeliness and accuracy, thus achieving real-time and accurate fire detection in forest areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of a lightweight smoke detection method based on self-evolution of an empirical model for forest fires provided by an embodiment of the present invention;

[0048] Figure 2 is a flow chart of generating a region to be confirmed according to an embodiment of the present invention;

[0049] Figure 3 This is a fire identification model diagram based on a surveillance camera provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0051] Reference Figure 1 The lightweight smoke detection method for forest fires based on the self-evolution of an empirical model provided by the embodiment of the present invention performs the following steps S1 to S7 for a target forest area to detect whether smoke is generated in the target forest area, thereby determining whether a fire has occurred:

[0052] Step S1. Real-time video of the target forest area is collected and converted into a time-sequential image sequence, with each frame of the image being divided into a preset number of image blocks;

[0053] The method of dividing each frame image into a preset number of image blocks in step S1 is as follows:

[0054] The size of each image in the image sequence is 256×256. For the image I at time t in the image sequence, t , divide it into image blocks of size 8×8, image I t The i-th image block in Contains 32×32 pixels, 1≤i≤64, Represents an image block The j-th pixel in , 1≤j≤1024, Respectively The red chroma color value, green chroma color value, and blue chroma color value.

[0055] Step S2. For the first image frame in the image sequence obtained in step S1, a pre-trained deep learning model is used that takes each image block in the image as input and uses as output a final identification corresponding to each image block in the image, indicating whether each image block contains smoke. The model determines whether each image block in the first image frame contains smoke, where a final identification of 1 indicates that the corresponding image block contains smoke, and a final identification of 0 indicates that the corresponding image block does not contain smoke.

[0056] Then, starting with the second frame in the image sequence, the following steps S3 to S7 are performed for each frame in sequence to detect whether smoke is generated in the target forest area:

[0057] Step S3: Using the statistical histogram method of the empirical sample data, a conditional probability model is constructed based on the R, G, and B of the image, with the forest background as the occurrence event E0, the smoke foreground as the occurrence event E1, and other foregrounds as the occurrence event E2;

[0058] An empirical model is constructed based on the conditional probability model. According to the empirical model, each pixel in the image is divided into smoke foreground pixels, forest background pixels, and other foreground pixels. Each image block of the image is traversed and the number of smoke foreground pixels, forest background pixels, and other foreground pixels in each image block is counted.

[0059] Based on the number of smoke foreground pixels, forest background pixels, and other foreground pixels in the image block, determine whether the image block contains smoke and match the preliminary identification corresponding to the image block. A preliminary identification of 1 indicates that the corresponding image block contains smoke, and a preliminary identification of 0 indicates that the corresponding image block does not contain smoke.

[0060] The specific steps of constructing the conditional probability model in step S3 are as follows:

[0061] Step S31. Construct conditional probability models based on the R, G, and B values ​​of the image, with the forest background as the occurrence event E0, the smoke foreground as the occurrence event E1, and other foregrounds as the occurrence event E2: P(E0|R), P(E1|R), P(E2|R), P(E0|G), P(E1|G), P(E2|G), P(E2|G), P(E1|B), P(E2|B), P(E2|B), where R represents the red chromaticity color value of the image, G represents the green chromaticity color value of the image, and B represents the blue chromaticity color value of the image;

[0062] Step S32: Construct an empirical model based on the conditional probability model, and classify each pixel that meets the following conditions as a smoke foreground pixel:

[0063]

[0064] Each pixel that meets the following conditions is classified as a forest background pixel:

[0065]

[0066] Where α1, α2, and α3 represent the importance coefficients corresponding to the conditional probabilities of R, G, and B, respectively, and Th1 is the joint conditional probability threshold. Represents image blocks Forest background countermeasure factor, smoke foreground countermeasure factor;

[0067] Each pixel that does not meet the above conditions is divided into other pixel points;

[0068] Step S33: Counting image blocks The number of forest background pixels Number of smoke foreground pixels Other pixel numbers If the following conditions are met:

[0069]

[0070] Then judge the image block For the smoke foreground area, the image block Matching preliminary identification Represents an image block There is smoke, otherwise the image block Matching preliminary identification Represents an image block There is no smoke in the scene; where Th2 is the pixel threshold of the smoke scene;

[0071] Step S34: Repeat step S33, traversing each image block until all image blocks in the image are judged.

[0072] The multiplication of conditional joint probabilities is to fit the joint conditional probability P(E|R,G,B). In essence, P(E|R,G,B) can more accurately predict event E based on R,G,B, but the joint conditions of R,G,B are sparse, and it is difficult to find enough samples to build the corresponding model. If the conditions are combined, the value space is 256×256×256, and the joint conditional model is difficult to establish. Therefore, single conditional models are established separately, and then they are connected to fit. α1, α2, and α3 are coefficients that reflect the correlation between R,G,B and event E. Th1 is used to remove other non-smoke foregrounds. The above conditions are essentially the possibility of smoke foreground and forest background. Whichever one prevails should be the one. If there is an error, it should be adjusted accordingly. Realize the self-evolution of the above empirical model.

[0073] Step S4. For each image block in the image, the preliminary identification is compared with the final identification of the image block at the same position in the adjacent previous image frame. If the preliminary identification of the subsequent image block in the two image blocks is different from the final identification of the previous image block, the subsequent image block is determined to be a pending confirmation area, and the process proceeds to step S5. Otherwise, the image block is determined to be a confirmed area, and the process proceeds to step S6 for the confirmed area.

[0074] Reference Figure 2 The method for determining the area to be confirmed in step S4 is as follows:

[0075]

[0076] Where, represents a modulo 2 addition operation, Represents the final identification of the i-th image block in the image at time t-1, if It means that the image block identification of the two frames before and after is different, that is, the image block is judged For the area to be confirmed, if It means that the image block identifiers of the two frames before and after are the same, that is, the image block is judged To confirm the area, For image blocks The final identification.

[0077] Step S5. Input the image blocks of the area to be confirmed into the pre-trained deep learning model, output the final identification of the image blocks of the area to be confirmed, and based on the preliminary identification and final identification of the image blocks of the area to be confirmed, evolve and update the empirical model, traverse all image blocks of the area to be confirmed, complete the self-evolution of the empirical model, and then proceed to step S7;

[0078] The specific steps of step S5 are as follows:

[0079] Step S51. If the image block For areas to be confirmed, a pre-trained deep learning model is used to determine the image block Does it contain smoke? If the image block If there is smoke, If the image block If there is no smoke,

[0080] Step S52: Image blocks of the area to be confirmed Based on preliminary identification Final logo The evolutionary update of the empirical model is as follows:

[0081]

[0082]

[0083] Where Δ is the evolution factor, Represents image blocks The forest background confrontation factor and the smoke foreground confrontation factor at the next moment;

[0084] Step S53: Repeat step S52 until all image blocks in the area to be confirmed are traversed and the self-evolution of the empirical model is completed.

[0085] Step S6. For each image block of the confirmed area, output the preliminary identification as the final identification, and then proceed to step S7;

[0086] Step S7: Based on the accumulation of judgment results of each image block, a pre-trained deep learning model is used to further verify each image block.

[0087] The specific method of further checking each image block using the pre-trained deep learning model in step S7 is as follows:

[0088] If the image I at time t in the image sequence t If there is no image block in the area to be confirmed, it is considered that the empirical model judgment has not changed, and the number of judgments of the empirical model CR increases by 1, otherwise CR increases by 0;

[0089] If any are all 0, 1≤i≤64, then when CR=N, where N is the preset number of judgments, and N is a positive integer, the number of smoke foreground pixels is selected For the image block with the most, a pre-trained deep learning model is used to determine whether the image block contains smoke, and the method described in step S5 is used to perform an evolutionary update of the empirical model;

[0090] If exists is 1, 1≤i≤64, then when CR=2N, select the number of smoke foreground pixels The least image block uses a pre-trained deep learning model to determine whether the image block contains smoke, and the method described in step S5 is used to evolve and update the empirical model.

[0091] The above steps are to prevent Keep it unchanged for a long time to prevent missed detection, and use sampling method to make deep learning judgments. If all are 0, it means there is no smoke for a long time. If the detection is missed, the consequences will be serious. Therefore, the number of judgments CR=N is set to sample and detect as soon as possible.

[0092] Reference Figure 3In one embodiment, a camera, hardware devices, and an alarm that are interconnected in sequence are used to implement the lightweight smoke detection method for self-evolution of an empirical model for forest fires provided by an embodiment of the present invention, wherein the camera is used to collect video of the target forest area in real time, and the hardware device includes a video image preprocessing area and a neural network processing area. The video image preprocessing area is used to complete the conversion of the video collected by the camera into a time-sequential image sequence as described in step S1, and divide each frame image into a preset number of image blocks. The neural network processing area is used to complete the detection of whether smoke is generated in the target forest area as described in steps S2 to S7. The alarm receives the smoke detection result output by the neural network processing area and issues an alarm if it is determined that smoke is generated.

[0093] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the scope of knowledge possessed by ordinary technicians in this field without departing from the spirit of the present invention.

Claims

1. A lightweight smoke detection method based on self-evolution of empirical models for forest fires, characterized by: For the target forest area, the following steps S1 to S7 are executed to detect whether smoke is generated in the target forest area, and then determine whether a fire has occurred: Step S1. Real-time video of the target forest area is captured and converted into a time-sequential image sequence. Each frame of the image is divided into a preset number of image blocks, and then the process proceeds to step S2. Step S2. For the first image frame in the image sequence obtained in step S1, a pre-trained deep learning model is used that takes each image block in the image as input and uses as output a final identification corresponding to each image block in the image, indicating whether each image block contains smoke. The model determines whether each image block in the first image frame contains smoke, where a final identification of 1 indicates that the corresponding image block contains smoke, and a final identification of 0 indicates that the corresponding image block does not contain smoke. Then, starting with the second frame in the image sequence, the following steps S3 to S7 are performed for each frame in sequence to detect whether smoke is generated in the target forest area: Step S3: Using the statistical histogram method of the empirical sample data, a conditional probability model is constructed based on the R, G, and B of the image, with the forest background as the occurrence event E0, the smoke foreground as the occurrence event E1, and other foregrounds as the occurrence event E2; An empirical model is constructed based on the conditional probability model. According to the empirical model, each pixel in the image is divided into smoke foreground pixels, forest background pixels, and other foreground pixels. Each image block of the image is traversed and the number of smoke foreground pixels, forest background pixels, and other foreground pixels in each image block is counted. Based on the number of smoke foreground pixels, forest background pixels, and other foreground pixels in the image block, determine whether the image block contains smoke and match the preliminary identification corresponding to the image block. A preliminary identification of 1 indicates that the corresponding image block contains smoke, and a preliminary identification of 0 indicates that the corresponding image block does not contain smoke. The specific steps of constructing the conditional probability model in step S3 are as follows: Step S31. Construct conditional probability models based on the R, G, and B values ​​of the image, with the forest background as the occurrence event E0, the smoke foreground as the occurrence event E1, and other foregrounds as the occurrence event E2: P(E0|R), P(E1|R), P(E2|R), P(E0|G), P(E1|G), P(E2|G), P(E2|G), P(E1|B), P(E2|B), P(E2|B), where R represents the red chromaticity color value of the image, G represents the green chromaticity color value of the image, and B represents the blue chromaticity color value of the image; Step S32: Construct an empirical model based on the conditional probability model, and classify each pixel that meets the following conditions as a smoke foreground pixel: Each pixel that meets the following conditions is classified as a forest background pixel: Where α1, α2, and α3 represent the importance coefficients corresponding to the conditional probabilities of R, G, and B, respectively, and Th1 is the joint conditional probability threshold. Represents image blocks Forest background countermeasure factor, smoke foreground countermeasure factor; Each pixel that does not meet the above conditions is divided into other pixel points; Step S33: Counting image blocks The number of forest background pixels Number of smoke foreground pixels Other pixel numbers If the following conditions are met: Then judge the image block For the smoke foreground area, the image block Matching preliminary identification Represents an image block There is smoke, otherwise the image block Matching preliminary identification Represents an image block There is no smoke in the scene; where Th2 is the pixel threshold of the smoke scene; Step S34. Repeat step S33, traversing each image block until all image blocks in the image are judged; Step S4. For each image block in the image, the preliminary identification is compared with the final identification of the image block at the same position in the adjacent previous image frame. If the preliminary identification of the subsequent image block in the two image blocks is different from the final identification of the previous image block, the subsequent image block is determined to be a pending confirmation area, and the process proceeds to step S5. Otherwise, the image block is determined to be a confirmed area, and the process proceeds to step S6 for the confirmed area. Step S5. Input the image blocks of the area to be confirmed into the pre-trained deep learning model, output the final identification of the image blocks of the area to be confirmed, and based on the preliminary identification and final identification of the image blocks of the area to be confirmed, evolve and update the empirical model, traverse all image blocks of the area to be confirmed, complete the self-evolution of the empirical model, and then proceed to step S7; The specific steps of step S5 are as follows: Step S51. If the image block For areas to be confirmed, a pre-trained deep learning model is used to determine the image block Does it contain smoke? If the image block If there is smoke, If the image block is judged If there is no smoke, Step S52: Image blocks of the area to be confirmed Based on preliminary identification Final logo The evolutionary update of the empirical model is as follows: Where Δ is the evolution factor, Represents image blocks The forest background confrontation factor and the smoke foreground confrontation factor at the next moment; Step S53: Repeat step S52 until all image blocks in the area to be confirmed are traversed and the self-evolution of the empirical model is completed; Step S6. For each image block of the confirmed area, output the preliminary identification as the final identification, and then proceed to step S7; Step S7: Based on the accumulation of judgment results of each image block, a pre-trained deep learning model is used to further verify each image block.

2. The lightweight smoke detection method based on the self-evolution of empirical models for forest fires according to claim 1 is characterized in that: The method of dividing each frame image into a preset number of image blocks in step S1 is as follows: The size of each image in the image sequence is 256×256. For the image I at time t in the image sequence, t , divide it into image blocks of size 8×8, image I t The i-th image block in Contains 32×32 pixels, 1≤i≤64, Represents an image block The j-th pixel in , 1≤j≤1024, Respectively The red chroma color value, green chroma color value, and blue chroma color value.

3. The lightweight smoke detection method based on the self-evolution of empirical models for forest fires according to claim 1 is characterized in that: The method for determining the area to be confirmed in step S4 is as follows: Where, represents a modulo 2 addition operation, Represents the final identification of the i-th image block in the image at time t-1, if It means that the image block identification of the two frames before and after is different, that is, the image block is judged For the area to be confirmed, if It means that the image block identifiers of the two frames before and after are the same, that is, the image block is judged To confirm the area, For image blocks The final identification.

4. The lightweight smoke detection method based on the self-evolution of empirical models for forest fires according to claim 1 is characterized in that: The specific method of further checking each image block using the pre-trained deep learning model in step S7 is as follows: If the image I at time t in the image sequence t If there is no image block in the area to be confirmed, it is considered that the empirical model judgment has not changed, and the number of judgments of the empirical model CR increases by 1, otherwise CR increases by 0; If any are all 0, 1≤i≤64, then when CR=N, where N is the preset number of judgments, and N is a positive integer, the number of smoke foreground pixels is selected For the image block with the most, a pre-trained deep learning model is used to determine whether the image block contains smoke, and the method described in step S5 is used to perform an evolutionary update of the empirical model; If exists is 1, 1≤i≤64, then when CR=2N, select the number of smoke foreground pixels The least image block uses a pre-trained deep learning model to determine whether the image block contains smoke, and the method described in step S5 is used to evolve and update the empirical model.

Images