Forest fire prevention method based on smoke recognition

Abstract

The invention belongs to the technical field of forest fire early warning, and discloses a forest fire prevention method based on smoke recognition, which comprises the following steps: acquiring heterogeneous image data from different monitoring devices, performing format unification and quality enhancement through a data preprocessing module, eliminating the difference of data from different sources, and obtaining preprocessed image data; a multi-scale feature extraction network based on deep learning is adopted, key features such as the color, texture and form of smoke are automatically learned and extracted from the preprocessed image data, and feature representation is obtained; according to the method, corresponding early warning threshold values are set according to different grades of fire judgment results, when the fire grade judgment results meet early warning conditions, early warning signals are triggered automatically, and early warning of the fire is achieved in the modes of message pushing, sound-light alarm and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of forest fire early warning, and in particular to a forest fire prevention method based on smoke recognition. Background Art

[0002] Smoke recognition is a key link in forest fire prevention. However, the surveillance image data obtained in the forest environment often come from various sources and formats. How to effectively fuse these heterogeneous data and extract key features is a major challenge. In addition, smoke presents variable colors and shapes under different environmental conditions, and traditional fixed threshold segmentation algorithms are difficult to accurately segment smoke areas of different concentrations. Even if the smoke area is identified through feature extraction and segmentation, it is still difficult to judge the severity of the fire. The correspondence between smoke concentration and fire level is vague, and there is a lack of quantitative estimation models. How to balance the contradictions between feature selection, segmentation algorithm and concentration estimation, improve the accuracy and reliability of smoke recognition, and achieve early warning and timely response to fire is a key technical problem that needs to be solved urgently. Summary of the invention

[0003] In order to solve the above technical problems, the present invention provides a forest fire prevention method based on smoke recognition, which can effectively solve key problems such as heterogeneous data processing, variable smoke recognition, and fire level judgment, and realizes full-process intelligent processing from heterogeneous image data collection to fire warning, thereby improving the accuracy of smoke recognition and the timeliness of fire warning.

[0004] A forest fire prevention method based on smoke recognition, the method comprising:

[0005] Obtain heterogeneous image data from different monitoring devices in a forest environment and preprocess them to obtain preprocessed image data;

[0006] A multi-scale feature extraction network based on deep learning is used to automatically learn and extract the key features of smoke from the pre-processed image data to obtain feature representation;

[0007] Aiming at the variable colors and shapes of smoke in different environments, an adaptive threshold segmentation algorithm is designed. By dynamically adjusting the segmentation threshold and processing the feature representation, the automatic segmentation of smoke areas with different concentrations is achieved to obtain the segmentation results.

[0008] According to the segmentation results, the key feature subset with the highest correlation with smoke concentration is screened out through the feature selection algorithm, and a smoke concentration estimation model is constructed to achieve the mapping from features to concentration and obtain the smoke concentration estimation result;

[0009] According to the smoke concentration estimation results, a corresponding relationship model between concentration and fire level is established to quantitatively evaluate the severity of the fire and obtain the fire level judgment result;

[0010] According to the fire situation judgment results of different levels, the corresponding warning threshold is set. When the fire situation level judgment result meets the warning conditions, the warning signal is automatically triggered.

[0011] Preferably, a multi-scale feature extraction network based on deep learning is used to automatically learn and extract key features of smoke from preprocessed image data, and the feature representation obtained includes:

[0012] Based on the deep learning algorithm, a multi-scale feature extraction network model is constructed;

[0013] For the input pre-processed image data, a multi-scale feature extraction network model is used to process it, automatically extracting the key features of multiple scales of color, texture and shape of the smoke area in the image;

[0014] Through the forward calculation process of the multi-scale feature extraction network model, a feature vector representing the smoke features in the input image is obtained;

[0015] According to the feature vector representing the smoke features in the input image, a smoke feature representation model is constructed;

[0016] Apply the constructed smoke feature representation model to the input image to determine whether there is a smoke area in the image. If there is a smoke area, determine the location and range of the smoke area according to the smoke feature representation model.

[0017] According to the smoke area position and range information determined by the smoke feature representation model, the corresponding smoke area image is segmented from the original input image;

[0018] The segmented smoke area image is input into the pre-trained smoke recognition model to determine the type and concentration attribute information of the smoke and obtain the final smoke detection result.

[0019] Preferably, in view of the variable colors and forms of smoke in different environments, an adaptive threshold segmentation algorithm is designed, and the segmentation threshold is dynamically adjusted to process the feature representation to achieve automatic segmentation of smoke areas with different concentrations. The segmentation results include:

[0020] Obtain images containing smoke areas of different densities, extract the color and texture features of the images, and obtain a smoke feature map;

[0021] The K-means clustering algorithm is used to divide the smoke feature map into multiple sub-areas, each of which corresponds to smoke of different concentrations;

[0022] By analyzing the local characteristics of each sub-region, the segmentation threshold of the corresponding region is adaptively determined;

[0023] Using the determined segmentation threshold, the sub-region is binarized to obtain the smoke segmentation result of the corresponding region;

[0024] Merge the segmentation results of each sub-region to obtain a complete smoke segmentation result map;

[0025] Perform morphological processing on the segmentation result image to obtain the final smoke segmentation result;

[0026] According to the smoke segmentation results, the concentration distribution of smoke in the image is judged to achieve automatic recognition and positioning of smoke areas with different concentrations.

[0027] Preferably, according to the segmentation results, a feature selection algorithm is used to select a subset of key features with the highest correlation with smoke concentration, a smoke concentration estimation model is constructed, and mapping from features to concentration is achieved. The smoke concentration estimation results obtained include:

[0028] Obtain smoke images, extract color, texture and area multi-dimensional feature parameters of the images, and construct feature sets;

[0029] The feature selection algorithm is used to calculate the correlation between each feature and smoke concentration, and the key feature subset with the highest correlation is screened out;

[0030] Based on the selected key feature subset, a smoke concentration estimation model is constructed, and the model is trained using a machine vision algorithm;

[0031] The trained smoke density estimation model is applied to the new smoke image to extract the key features of the image, and the smoke density estimation value of the image is obtained by mapping the features to the concentration.

[0032] Preferably, based on the smoke concentration estimation result, a corresponding relationship model between the concentration and the fire level is established to quantitatively evaluate the severity of the fire, and the fire level judgment result obtained includes:

[0033] According to the pre-established correspondence model between smoke density and fire level, the estimated value of smoke density at the current moment is obtained, and the initial value of the fire level to which it belongs is determined;

[0034] Acquire multiple frames of continuous smoke images, and calculate the area change rate and concentration change rate of the smoke region through image analysis as attribute features to describe the dynamic change information of the smoke;

[0035] The dynamic attribute characteristics of smoke duration, area change rate and concentration change rate are input into the fire level judgment model, and the quantitative value of the fire severity at the current moment is calculated through model reasoning;

[0036] The fire severity quantification value is compared with the preset fire level threshold. If it exceeds a certain level threshold, the current fire level is determined to be the level corresponding to the threshold.

[0037] The initial fire level and the fire level obtained by dynamic information analysis are comprehensively considered, and the final fire level judgment result is calculated through a weighted fusion algorithm;

[0038] During the continuous multi-frame image analysis process, if the smoke area and concentration increase sharply in a short period of time, it is judged that the fire is expanding rapidly and the fire alarm is immediately triggered.

[0039] Preferably, corresponding warning thresholds are set for different levels of fire situation judgment results. When the fire situation level judgment result meets the warning condition, the warning signal is automatically triggered, including:

[0040] According to the acquired fire monitoring data, a machine learning algorithm is used to judge the fire level and obtain the fire level judgment result;

[0041] According to different fire level judgment results, a warning threshold corresponding to the current fire level is determined from pre-set warning thresholds;

[0042] If the fire level judgment result reaches or exceeds the warning threshold, the warning signal will be automatically triggered;

[0043] According to the warning signal, a fire warning message is sent to the pre-set recipient by calling the message push interface;

[0044] According to the early warning signal, the connected sound and light alarm device is controlled to emit sound and flash to give an on-site fire warning;

[0045] Obtain fire situation judgment results and warning information, generate fire situation warning logs, and record information such as the time, location, and level of fire situation warnings;

[0046] The fire warning log is uploaded to the cloud server for storage, and the historical records and statistical information of the fire warning are displayed through a visual interface to provide data support for fire prevention and emergency decision-making.

[0047] Preferably, the method also includes: realizing visual display of smoke recognition results, and providing manual verification and feedback mechanism, continuously optimizing the algorithm model according to feedback information, and the visual display includes a comprehensive display of original image, segmentation results, concentration estimation and fire level information.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] The present invention discloses a forest fire prevention method and system based on smoke recognition. For heterogeneous image data from different monitoring devices, format unification and quality enhancement are achieved through preprocessing to eliminate data differences. A multi-scale feature extraction network is used to automatically learn the key features of smoke, and an adaptive threshold segmentation algorithm is designed to achieve accurate segmentation of smoke areas with different concentrations. Feature subsets highly correlated with smoke concentration are screened out through feature selection to construct a concentration estimation model. Combined with the dynamic change information of smoke, a corresponding relationship model between concentration and fire level is established to achieve quantitative evaluation of the severity of the fire. An early warning threshold is set according to the evaluation result, and an early warning signal is automatically triggered. The present invention also provides a visual display and manual feedback mechanism to continuously optimize the algorithm model. The present invention can effectively solve key problems such as heterogeneous data processing, multi-variable smoke recognition, and fire level judgment, and realizes the full-process intelligent processing from heterogeneous image data acquisition to fire early warning, thereby improving the accuracy of smoke recognition and the timeliness of fire early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0051] Figure 1 A schematic flow chart of a forest fire prevention method based on smoke recognition according to an embodiment of the present invention;

[0052] Figure 2 A schematic diagram of the smoke concentration estimation model structure according to an embodiment of the present invention;

[0053] Figure 3 Schematic diagram of the spatial pyramid pooling structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present disclosure should be understood by people with ordinary skills in the field to which the present disclosure belongs. The "first", "second" and similar words used in the embodiments of the present disclosure do not indicate any order, quantity or importance, but are only used to distinguish different components. "Including" or "comprising" and similar words mean that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0056] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] Embodiment 1

[0058] like Figure 1 As shown, the embodiment of the present invention provides a forest fire prevention method based on smoke recognition, which may specifically include:

[0059] Step S101, obtaining heterogeneous image data from different monitoring devices, unifying the format and enhancing the quality through a data preprocessing module, eliminating the differences between data from different sources, and obtaining preprocessed image data.

[0060] Obtain heterogeneous raw image data from different monitoring devices, and use data preprocessing technology to process image data of different formats and qualities. Use image format conversion algorithms to convert image data of different formats into a unified format to eliminate the impact of format differences. According to image quality assessment indicators, perform quality assessment on image data, and use image enhancement algorithms to process images with poor quality to improve image quality. Perform noise removal on image data, and remove noise interference in the image by combining wavelet transform and threshold processing to obtain a denoised image. Use image contrast enhancement technology to enhance the contrast of the image through algorithms such as histogram equalization to make the image details clearer. Perform geometric correction on the image, and use algorithms such as perspective transformation to correct the geometric distortion of the image to obtain a geometrically corrected image. Encapsulate the preprocessed image data in a unified format to obtain monitoring image data with a unified format and enhanced quality, providing high-quality data support for subsequent image analysis applications.

[0061] Step S102, using a multi-scale feature extraction network based on deep learning, automatically learns and extracts key features such as color, texture, and shape of smoke from the preprocessed image data to obtain feature representation.

[0062] According to the deep learning algorithm, a multi-scale feature extraction network model is constructed to automatically learn and extract smoke features. The pre-processed image data is processed by the constructed multi-scale feature extraction network model to automatically extract the key features of multiple scales such as color, texture and morphology of the smoke area in the image. Through the forward calculation process of the multi-scale feature extraction network model, the feature vector representing the smoke features in the input image is obtained. According to the extracted multi-scale features of smoke, a smoke feature representation model is constructed to characterize the overall features of the smoke area in the input image. The constructed smoke feature representation model is applied to the input image to determine whether there is a smoke area in the image. If there is a smoke area, the position and range of the smoke area are determined according to the smoke feature representation model. According to the smoke area position and range information determined by the smoke feature representation model, the corresponding smoke area image is segmented from the original input image. The segmented smoke area image is input into the pre-trained smoke recognition model to determine the attribute information such as the type and concentration of the smoke, and obtain the final smoke detection result.

[0063] Step S103, in view of the variable colors and forms of smoke in different environments, an adaptive threshold segmentation algorithm is designed, and the feature representation is processed by dynamically adjusting the segmentation threshold to achieve automatic segmentation of smoke areas of different concentrations and obtain segmentation results. The specific processing method includes multi-level threshold segmentation of the feature map and adaptively adjusting the threshold according to the local characteristics of the image.

[0064] An image containing smoke areas of different densities is obtained, and the color and texture features of the image are extracted to obtain a smoke feature map. According to the smoke feature map, the K-means clustering algorithm is used to divide the feature map into multiple sub-regions, each of which corresponds to smoke of different densities. For each sub-region, the segmentation threshold of the region is adaptively determined by analyzing its local characteristics, including color distribution, texture complexity, etc. The sub-region is binarized using the determined segmentation threshold to obtain the smoke segmentation result of the region. The segmentation results of each sub-region are merged to obtain a complete smoke segmentation result map. The segmentation result map is morphologically processed to eliminate small noise areas to obtain the final smoke segmentation result. According to the segmentation result, the concentration distribution of smoke in the image is judged to realize automatic recognition and positioning of smoke areas of different concentrations.

[0065] Step S104, based on the segmentation results, extract characteristic parameters such as color, texture, and area of ​​the smoke, select the key feature subset with the highest correlation with the smoke concentration through the feature selection algorithm, build a smoke concentration estimation model, realize the mapping from features to concentration, and obtain the smoke concentration estimation result.

[0066] Obtain a smoke image, extract multi-dimensional feature parameters such as color, texture, and area of ​​the image, and construct a feature set; adopt a feature selection algorithm to screen out the key feature subset with the highest correlation by calculating the correlation between each feature and the smoke concentration; construct a smoke concentration estimation model based on the screened key feature subset, and use a machine vision algorithm to train the model; during the model training process, if the estimation error of the current model is greater than the preset threshold, return to step 3 and readjust the model parameters until the model converges; apply the trained smoke concentration estimation model to a new smoke image, extract the key features of the image, and obtain the smoke concentration estimation value of the image through mapping from features to concentration; estimate the smoke concentration of a series of smoke images obtained continuously, and obtain the changing trend of the smoke concentration over a period of time; based on the changing trend of the smoke concentration, determine whether the current smoke concentration exceeds the warning threshold, and if so, issue a warning message to prompt the user to take corresponding measures.

[0067] Specific implementation process:

[0068] The smoke concentration estimation model is as follows: Figure 2 As shown in the figure, the overall system can be divided into three modules: input module, feature extraction module and output module. (1) The input module consists of a smoke slice image to be tested and multiple benchmark concentration sample images. The benchmark concentration samples come from the constructed training data set Part2 (2874 background slice images and 12370 smoke slice images). 100 smoke images of different concentrations are extracted as reference benchmarks, with a specific concentration range of 0.00 to 1.00. (2) The feature extraction module contains a pair of sub-networks, which are used to extract the feature vectors of the smoke image to be tested and the benchmark concentration sample image. Since the differences between them need to be compared through these two feature vectors, sub-network 1 and sub-network 2 should have the same structure and share weights to ensure that the feature vectors they output are in the same dimensional space. (3) The output module calculates the Euclidean distance between the feature vector of the smoke to be tested and the feature vector of each benchmark sample, and selects the concentration corresponding to the benchmark sample with the highest similarity from the 100 samples of different concentrations as the concentration value of the smoke image to be tested.

[0069] In order to ensure that the sub-network can still maintain high accuracy in the deep network, the ResNet50 structure is introduced as the basis of the sub-network. ResNet is a network structure that adds residual connections and skip layer connection structures on the basis of VGG. It will not degrade due to the depth of the network, so as to stabilize the accuracy of the twin network. Since the smoke images that need concentration estimation are usually intercepted from the segmented smoke area, these images are of different sizes, and the fully connected layer in the Resnet50 network requires that the input sample size must be the same. Therefore, the twin network will resize the input samples for preprocessing, which will cause the image to be compressed or interpolated, thereby losing some smoke concentration features, and ultimately affecting the accuracy of the network output results. To solve this problem, the spatial pyramid pooling technology can be introduced to enable the network to process smoke images of different sizes.

[0070] The spatial pyramid pooling structure is as follows Figure 3 , this structure can be added before the fully connected layer to perform special operations on the feature map output by the convolution layer. Assuming that the output of the convolution layer is a feature with height h, width w, and channel c, take w / 4 and h / 4 as the width and height of the pooling for pooling on each channel of the feature, and then take half of the width and height of the feature as the width and height of the pooling for pooling, and finally take the original width and height of the feature as the width and height of the pooling for pooling. The vectors obtained by the above operations can finally be combined into a 21-dimensional vector. This fixed result can be obtained by inputting images of any size, thus solving the problem of smoke concentration information loss.

[0071] Step S105: Based on the smoke density estimation result, a corresponding relationship model between the density and the fire level is established, and the dynamic information such as the duration and change trend of the smoke is comprehensively considered to quantitatively evaluate the severity of the fire and obtain the fire level judgment result. The dynamic information is obtained by analyzing continuous multi-frame images, including the smoke area change rate, concentration change rate, etc.

[0072] According to the pre-established correspondence model between smoke density and fire level, the estimated smoke density at the current moment is obtained to determine the initial value of the fire level to which it belongs. Continuous multi-frame smoke images are obtained, and the area change rate and concentration change rate of the smoke area are calculated through image analysis as attribute features describing the dynamic change information of smoke. The dynamic attribute features such as smoke duration, area change rate and concentration change rate are input into the fire level judgment model, and the quantitative value of the fire severity at the current moment is calculated through model reasoning. The quantitative value of the fire severity is compared with the preset fire level threshold. If it exceeds a certain level threshold, the current fire level is judged to be the level corresponding to the threshold. The initial fire level and the fire level obtained by dynamic information analysis are comprehensively considered, and the final fire level judgment result is calculated by weighted fusion algorithm. In the process of continuous multi-frame image analysis, if the smoke area and concentration rise sharply in a short period of time, it is judged that the fire is expanding rapidly and the fire alarm is triggered immediately. For large smoke areas, they are divided into multiple sub-areas by segmentation algorithm, and the fire level of each sub-area is judged separately to improve the precision and accuracy of analysis.

[0073] Dynamic attribute characteristics such as smoke duration, area change rate and concentration change rate are input into the fire level judgment model, and the quantitative value of the fire severity at the current moment is calculated through model reasoning.

[0074] The dynamic attribute characteristic data of smoke, such as duration, area change rate and concentration change rate, are obtained and used as the input of the fire level judgment model. According to the pre-established fire level judgment model, the machine learning algorithm is used to train and learn the input dynamic attribute characteristic data to obtain a trained fire level judgment model. The dynamic attribute characteristic data of smoke obtained at the current moment is input into the trained fire level judgment model, and the quantitative value of the fire severity at the current moment is obtained through model reasoning and calculation. The fire severity at the current moment is judged by comparing the quantitative value of the fire severity with the preset fire level threshold. If the quantitative value of the fire severity at the current moment exceeds the preset fire level threshold, the warning information is triggered to notify relevant personnel to take emergency measures. The dynamic attribute characteristic data of smoke is continuously obtained, and the data is input into the fire level judgment model in real time for reasoning and calculation, so as to dynamically evaluate and predict the development trend of the fire. According to the dynamic evaluation and prediction results of the fire, the emergency plan and resource scheduling are dynamically adjusted to ensure that corresponding response measures are taken at different stages of the fire development.

[0075] Step S106, for different levels of fire judgment results, set corresponding warning thresholds. When the fire level judgment result meets the warning conditions, the warning signal is automatically triggered, and early warning of the fire is achieved through message push, sound and light alarm, etc.

[0076] According to the acquired fire monitoring data, the machine learning algorithm is used to judge the fire level and obtain the fire level judgment result. For different fire level judgment results, the warning threshold corresponding to the current fire level is determined from the pre-set warning threshold. If the fire level judgment result reaches or exceeds the warning threshold, the warning signal is automatically triggered. According to the warning signal, the fire warning message is sent to the pre-set recipient by calling the message push interface. According to the warning signal, the sound and light alarm device connected is controlled to emit sound and flash to warn the on-site fire. The fire judgment result and warning information are obtained, and a fire warning log is generated to record the time, location, level and other information of the fire warning. The fire warning log is uploaded to the cloud server for storage, and the historical records and statistical information of the fire warning are displayed through a visual interface to provide data support for fire prevention and emergency decision-making.

[0077] Step S107, realize the visualization display of smoke recognition results, and provide manual verification and feedback mechanism, continuously optimize the algorithm model according to the feedback information, and improve the recognition accuracy and reliability. The visualization display includes the comprehensive display of the original image, segmentation results, concentration estimation and fire level information.

[0078] Obtain the original image to be identified, and preprocess the original image, including image enhancement and noise removal, to improve image quality. Use a semantic segmentation algorithm based on deep learning to segment the smoke region of the preprocessed image to obtain a segmentation result map. According to the segmentation result map, extract the features of the smoke region, including smoke color, texture, motion and other features, and obtain the smoke concentration estimation value through feature fusion. According to the smoke concentration estimation value, combined with the preset fire level threshold, judge the current fire level and obtain the fire level result. Visualize and display the original image, segmentation result map, concentration estimation value and fire level result to generate an intuitive recognition result display interface. Provide manual verification function, allowing users to annotate and feedback the recognition results, use the feedback results as training data, and continuously optimize the smoke recognition model. Adopt an active learning strategy, select the most valuable samples for users to annotate according to the uncertainty of model prediction, improve the efficiency of feedback data, and accelerate the model optimization process.

[0079] The embodiments of the present disclosure are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present disclosure should be included in the scope of protection of the present disclosure.

Claims

1. A forest fire prevention method based on smoke recognition, characterized in that: The method comprises: Obtain heterogeneous image data from different monitoring devices in a forest environment and preprocess them to obtain preprocessed image data; A multi-scale feature extraction network based on deep learning is used to automatically learn and extract the key features of smoke from the pre-processed image data to obtain feature representation; Aiming at the variable colors and shapes of smoke in different environments, an adaptive threshold segmentation algorithm is designed. By dynamically adjusting the segmentation threshold and processing the feature representation, the automatic segmentation of smoke areas with different concentrations is achieved to obtain the segmentation results. According to the segmentation results, the key feature subset with the highest correlation with smoke concentration is screened out through feature selection algorithm, and a smoke concentration estimation model is constructed to achieve the mapping from features to concentration and obtain the smoke concentration estimation result; According to the smoke concentration estimation results, a corresponding relationship model between concentration and fire level is established to quantitatively evaluate the severity of the fire and obtain the fire level judgment result; According to the fire situation judgment results of different levels, the corresponding warning threshold is set. When the fire situation level judgment result meets the warning conditions, the warning signal is automatically triggered.

2. The method according to claim 1, characterized in that A multi-scale feature extraction network based on deep learning is used to automatically learn and extract the key features of smoke from the preprocessed image data. The feature representations obtained include: Based on the deep learning algorithm, a multi-scale feature extraction network model is constructed; For the input pre-processed image data, a multi-scale feature extraction network model is used to process it, automatically extracting the key features of multiple scales of color, texture and shape of the smoke area in the image; Through the forward calculation process of the multi-scale feature extraction network model, a feature vector representing the smoke features in the input image is obtained; According to the feature vector representing the smoke features in the input image, a smoke feature representation model is constructed; Apply the constructed smoke feature representation model to the input image to determine whether there is a smoke area in the image. If there is a smoke area, determine the location and range of the smoke area according to the smoke feature representation model. According to the smoke area position and range information determined by the smoke feature representation model, the corresponding smoke area image is segmented from the original input image; The segmented smoke area image is input into the pre-trained smoke recognition model to determine the type and concentration attribute information of the smoke and obtain the final smoke detection result.

3. The method according to claim 1, characterized in that Aiming at the changeable colors and forms of smoke in different environments, an adaptive threshold segmentation algorithm is designed. By dynamically adjusting the segmentation threshold and processing the feature representation, the automatic segmentation of smoke areas with different concentrations is achieved. The segmentation results include: Obtain images containing smoke areas of different densities, extract the color and texture features of the images, and obtain a smoke feature map; The K-means clustering algorithm is used to divide the smoke feature map into multiple sub-areas, each of which corresponds to smoke of different concentrations; By analyzing the local characteristics of each sub-region, the segmentation threshold of the corresponding region is adaptively determined; Using the determined segmentation threshold, the sub-region is binarized to obtain the smoke segmentation result of the corresponding region; Merge the segmentation results of each sub-region to obtain a complete smoke segmentation result map; Perform morphological processing on the segmentation result image to obtain the final smoke segmentation result; According to the smoke segmentation results, the concentration distribution of smoke in the image is judged to achieve automatic recognition and positioning of smoke areas with different concentrations.

4. The method according to claim 1, characterized in that: According to the segmentation results, the key feature subset with the highest correlation with smoke concentration is screened out through the feature selection algorithm, and a smoke concentration estimation model is constructed to achieve the mapping from features to concentration. The smoke concentration estimation results obtained include: Obtain smoke images, extract color, texture and area multi-dimensional feature parameters of the images, and construct feature sets; The feature selection algorithm is used to calculate the correlation between each feature and smoke concentration, and the key feature subset with the highest correlation is screened out; Based on the selected key feature subset, a smoke concentration estimation model is constructed, and the model is trained using a machine vision algorithm; The trained smoke density estimation model is applied to the new smoke image to extract the key features of the image, and the smoke density estimation value of the image is obtained by mapping the features to the concentration.

5. The method according to claim 1, characterized in that According to the smoke concentration estimation results, a corresponding relationship model between concentration and fire level is established to quantitatively evaluate the severity of the fire. The fire level judgment results include: According to the pre-established correspondence model between smoke density and fire level, the estimated value of smoke density at the current moment is obtained, and the initial value of the fire level to which it belongs is determined; Acquire multiple frames of continuous smoke images, and calculate the area change rate and concentration change rate of the smoke region through image analysis as attribute features to describe the dynamic change information of the smoke; The dynamic attribute characteristics of smoke duration, area change rate and concentration change rate are input into the fire level judgment model, and the quantitative value of the fire severity at the current moment is calculated through model reasoning; The fire severity quantification value is compared with the preset fire level threshold. If it exceeds a certain level threshold, the current fire level is determined to be the level corresponding to the threshold. The initial fire level and the fire level obtained by dynamic information analysis are comprehensively considered, and the final fire level judgment result is calculated through a weighted fusion algorithm; During the continuous multi-frame image analysis process, if the smoke area and concentration increase sharply in a short period of time, it is judged that the fire is expanding rapidly and the fire alarm is immediately triggered.

6. The method according to claim 1, characterized in that According to the fire situation judgment results of different levels, the corresponding warning threshold is set. When the fire situation level judgment result meets the warning conditions, the warning signal is automatically triggered, including: According to the acquired fire monitoring data, a machine learning algorithm is used to judge the fire level and obtain the fire level judgment result; According to different fire level judgment results, a warning threshold corresponding to the current fire level is determined from pre-set warning thresholds; If the fire level judgment result reaches or exceeds the warning threshold, the warning signal will be automatically triggered; According to the warning signal, a fire warning message is sent to the pre-set recipient by calling the message push interface; According to the early warning signal, the connected sound and light alarm device is controlled to emit sound and flash to give an on-site fire warning; Obtain fire situation judgment results and warning information, generate fire situation warning logs, and record information such as the time, location, and level of fire situation warnings; The fire warning log is uploaded to the cloud server for storage, and the historical records and statistical information of the fire warning are displayed through a visual interface to provide data support for fire prevention and emergency decision-making.

7. The method according to claim 1, characterized in that The method also includes: realizing visual display of smoke recognition results, providing manual verification and feedback mechanism, and continuously optimizing the algorithm model according to feedback information. The visual display includes a comprehensive display of original images, segmentation results, concentration estimation and fire level information.

Images