A smoke density estimation method based on long- and short-range feature fusion

Abstract

The present invention proposes a smoke density estimation method by fusing long- and short-range features, the method comprising: inputting a color smoke composite image into a content extraction module and encoding it in combination with a lightweight network to obtain basic features of the smoke; inputting the basic features of the smoke into an information decoder for decoding to obtain smoke density maps of different levels; processing the smoke density map of the current layer to obtain a fine smoke density map; reconstructing the fine smoke density map and fusing it with the fine smoke density map to obtain a credible density map as the final output; the present invention calculates the position of smoke pixels from the color smoke monitoring image and infers the concentration, which is expected to provide more credible judgment clues for early fire warning, fire prediction and even rough classification of combustible materials.

Description

Technical Field

[0001] The present invention relates to the fields of deep learning and computer vision, and in particular to a smoke concentration estimation method by fusing long- and short-range features. Background Art

[0002] Early prevention and detection of fire has always been a topic of great research significance. Accurate, effective and timely smoke detection is an important means to control the spread of fire. However, complex scenes put forward higher requirements for smoke detection. In order to provide more clues, this invention targets the smoke concentration estimation task, which is a more accurate smoke detection subtask.

[0003] Long-distance global features can capture the overall distribution and shape of smoke in the scene, while short-distance local features can extract detailed information about the smoke. Effectively extracting and fusing long-distance dependency information and short-distance texture features can enhance the model's ability to express diverse smoke forms. However, illumination changes and background interference in complex environments make existing detection methods susceptible to external interference during the reasoning process, resulting in greater uncertainty in the detection results and affecting the reliability of detection. Summary of the invention

[0004] In view of the above situation, the main purpose of the present invention is to propose a smoke concentration estimation method by fusing long- and short-range features to solve the above technical problems.

[0005] The present invention proposes a smoke concentration estimation method based on long- and short-range feature fusion, the method comprising the following steps:

[0006] Step 1: construct a content extraction module based on depthwise separable convolution, and set up several layers of long and short distance feature modules in a stacked manner to construct a decoder;

[0007] Step 2: Input the color smoke composite image into the content extraction module and encode it with the lightweight network to obtain the mid-level smoke features and the high-level smoke features respectively. Downsample the mid-level smoke features and concatenate them with the high-level smoke features to obtain the basic features of the smoke.

[0008] Step 3: Input the basic features of the smoke into the information decoder, and use the long-distance and short-distance feature modules to decode the basic features of the smoke, and obtain the long-distance features and short-distance features of the smoke respectively;

[0009] The long-distance features, short-distance features and basic features of smoke are fused to obtain the smoke density map of the current layer;

[0010] The smoke density map of the current layer is used as the input of the long-short distance feature module of the next layer. After passing through several layers of long-short distance feature modules, smoke density maps of different levels are obtained.

[0011] Step 4: perform spatial supplementation and upsampling operations on the channel information in the smoke density map of the first layer in sequence to obtain an upsampling feature map;

[0012] Perform bilinear interpolation and upsampling operations on smoke density maps of different layers except the first layer in sequence to obtain upsampled density maps;

[0013] Perform a convolution operation on the upsampled feature map to obtain the convolution result of the upsampled feature map, perform a convolution operation and a global average pooling operation on the upsampled concentration map in sequence to obtain the average pooling result of the upsampled concentration map, and add the convolution result of the upsampled feature map and the average pooling result to obtain a multi-scale feature group;

[0014] The multi-scale feature group is input into the head module of the concentration map refinement module for splicing operation to obtain a fine smoke concentration map;

[0015] Step 5: reconstruct the smoke fine concentration map to obtain a reconstructed concentration map;

[0016] The reconstructed concentration map is fused with the smoke fine concentration map to obtain the final output credible concentration map.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] 1. The present invention realizes global dependency calculation through lightweight attention based on pooling operation, and introduces learnable fusion coefficient to perform adaptive global, local and original smoke feature fusion, so as to realize effective fusion of multiple features for smoke;

[0019] 2. The present invention introduces global and local features in the decoding process through a concentration information decoder based on long and short distance feature modules, thereby achieving more accurate concentration information calculation and smoke pixel positioning;

[0020] 3. The present invention proposes a Bayesian credible module to reconstruct the concentration map, thereby approximating the point estimate calculated by the concentration map, and then using the uncertainty of the reconstruction module to approximate the uncertainty of the previous step prediction, thereby achieving joint updating of reconstruction parameters and uncertain hyperparameters to reduce uncertainty, and ultimately improving the credibility of the concentration estimate.

[0021] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description or learned through embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1This is a flow chart of the steps of the smoke concentration estimation method based on long- and short-range feature fusion proposed by the present invention.

[0023] Figure 2 This is a structural diagram of the smoke concentration estimation method based on long- and short-range feature fusion proposed in the present invention.

[0024] Figure 3 This is a structural diagram of the long- and short-distance feature basic module of the smoke concentration estimation method based on the long- and short-distance feature fusion proposed in the present invention.

[0025] Figure 4 This is a structural diagram of the concentration map refinement module of the smoke concentration estimation method based on long- and short-range feature fusion proposed in the present invention.

[0026] Figure 5 This is the workflow of the Bayesian credible module of the smoke concentration estimation method based on long- and short-range feature fusion proposed by the present invention.

[0027] Figure 6 This is a comparison chart of the prediction results of different smoke segmentation or concentration estimation methods for the smoke concentration estimation method based on long- and short-range feature fusion proposed in the present invention. DETAILED DESCRIPTION

[0028] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout are the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0029] These and other aspects of the embodiments of the present invention will be apparent with reference to the following description and accompanying drawings. In these descriptions and accompanying drawings, some specific implementations of the embodiments of the present invention are specifically disclosed to provide some ways to implement the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0030] See also Figure 1 The embodiment of the present invention provides a smoke density estimation method by integrating long- and short-range features, the method comprising the following steps:

[0031] Step 1: Construct a content extraction module based on depthwise separable convolution, and set up several layers of long and short distance feature modules in a stacked manner to construct a decoder.

[0032] Step 2: Input the color smoke composite image into the content extraction module and encode it with a lightweight network to obtain mid-level smoke features and high-level smoke features respectively. Downsample the mid-level smoke features and concatenate them with the high-level smoke features to obtain the basic features of the smoke.

[0033] See also Figure 2 In step 2, the color smoke composite image is input into the content extraction module and encoded with the lightweight network to obtain the mid-level smoke features and the high-level smoke features respectively. The mid-level smoke features are downsampled and concatenated with the high-level smoke features to obtain the basic features of the smoke. The corresponding relationship in the process is as follows:

[0034] ;

[0035] in, Indicates the basic characteristics of smoke, represents the splicing operation along the channel direction, represents the downsampling operation based on bilinear interpolation, Indicates the characteristics of mid-level smoke. Represents high-level smoke characteristics.

[0036] Furthermore, a lightweight network with a total of 28 layers is used as the backbone network. The 11th layer of the network outputs mid-level smoke features, and the 28th layer outputs high-level smoke features. Both features come from the common basic backbone network, so the combined features of the two are called basic features.

[0037] Step 3: Input the basic features of the smoke into the information decoder, in which several layers of long and short distance feature modules are stacked;

[0038] The basic features of smoke are decoded using the long- and short-range feature modules to obtain the long-range features and short-range features of smoke respectively.

[0039] The long-distance features, short-distance features and basic features of smoke are fused to obtain the smoke density map of the current layer;

[0040] The smoke density map of the current layer is used as the input of the long-short distance feature module of the next layer to obtain smoke density maps of different levels.

[0041] See also Figure 3 In step 3, the basic features of the smoke are decoded using the long- and short-distance feature modules to obtain the long-distance features and short-distance features of the smoke, respectively, which specifically includes the following sub-steps:

[0042] Inputting the basic features of the smoke into the information decoder, performing average pooling operations and convolution operations in sequence, to obtain calculated components, wherein the calculated components include a first component, a second component, and a third component;

[0043] The second component and the third component are subjected to matrix multiplication and normalization in sequence to obtain the extracted long-distance dependency;

[0044] The extracted long-distance dependencies are subjected to random neuron discarding operations and matrix multiplication operations combined with the first component in sequence to obtain the long-distance features of the smoke;

[0045] The basic features of smoke are input into the information decoder to perform group convolution operation, normalization operation along the channel direction and point-by-point convolution operation in sequence to obtain the intermediate local features;

[0046] The intermediate local features are activated, convolved point by point, and randomly discarded neurons in sequence to obtain the short-range features of the smoke.

[0047] In the step of inputting the basic features of smoke into the information decoder and performing average pooling and convolution operations in sequence to obtain the calculated components, the corresponding relationship in the process is as follows:

[0048] ;

[0049] in, Indicates The calculated components, Indicates convolution operations, Indicates Average pooling operation;

[0050] In the step of performing matrix multiplication and normalization on the second component and the third component in sequence to obtain the extracted long-distance dependency, the corresponding relationship in the process is as follows:

[0051] ;

[0052] in, represents the extracted long-distance dependency, represents normalization processing, represents a channel-by-channel multiplication operation, represents division, Indicates the square root of the number of channels that are dependent on long distances. represents the second component, represents the third component;

[0053] It should be noted that Used to prevent overflow in multiplication operations.

[0054] In the step of performing random discard neuron operation and matrix multiplication operation combined with the first component on the extracted long-distance dependencies to obtain the long-distance features of the smoke, the corresponding relationship in the process is as follows:

[0055] ;

[0056] in, Indicates the long-distance characteristics of smoke, represents the random discarding of neuron operations, represents the first component;

[0057] In the step of inputting the basic features of smoke into the information decoder and performing group convolution operation, normalization operation along the channel direction and point-by-point convolution operation in sequence to obtain the intermediate local features, the corresponding relationship in the process is as follows:

[0058] ;

[0059] in, represents the intermediate local features, represents the point-by-point convolution operation of the first 1×1 window, represents the normalization operation along the channel direction, Represents a grouped convolution operation of a 3×3 window;

[0060] In the steps of performing activation operations, point-by-point convolution operations, and random neuron discarding operations on the intermediate local features in order to obtain the short-range features of smoke, the corresponding relationship between the processes is as follows:

[0061] ;

[0062] in, Indicates the short-range characteristics of smoke, represents the point-by-point convolution operation of the second 1×1 window, Represents a GeLU activation operation.

[0063] Furthermore, the first long-short distance feature module accepts basic features as input, and the present invention stacks the long-short distance feature modules 4 times, and the fourth group of long-short distance fused semantic features is a collection of 4 groups of long-short distance fused semantic features.

[0064] In the step of fusing the long-distance features, short-distance features and basic features of smoke to obtain the smoke density map of the current layer, the corresponding relationship between the process is as follows:

[0065] ;

[0066] in, Represents the smoke density map of the current layer, The fusion coefficient representing the long-distance characteristics of smoke, The fusion coefficient representing the short-range characteristics of smoke, The fusion coefficient representing the basic characteristics of smoke, represents the normalized coefficient, represents the indicator function, Indicates Coefficient .

[0067] In the step of using the smoke density map of the current layer as the input of the long-short distance feature module of the next layer to obtain smoke density maps of different levels, the corresponding relationship between the process is as follows:

[0068] ;

[0069] in, Indicates Smoke density diagram of long and short distance feature modules, Indicates Smoke density diagram of long and short distance feature modules, Indicates Long and short distance feature modules.

[0070] Step 4: perform spatial supplementation and upsampling operations on the channel information in the smoke density map of the first layer in sequence to obtain an upsampling feature map;

[0071] Perform bilinear interpolation and upsampling operations on smoke density maps of different layers except the first layer in sequence to obtain upsampled density maps;

[0072] Perform a convolution operation on the upsampled feature map to obtain the convolution result of the upsampled feature map, perform a convolution operation and a global average pooling operation on the upsampled concentration map in sequence to obtain the average pooling result of the upsampled concentration map, and add the convolution result of the upsampled feature map and the average pooling result to obtain a multi-scale feature group;

[0073] The multi-scale feature group is input into the head module of the concentration map refinement module for splicing operation to obtain the smoke fine concentration map.

[0074] See also Figure 4 In step 4, the channel information in the smoke density map of the first layer is sequentially subjected to spatial supplementation and upsampling operations to obtain an upsampling feature map. The corresponding relationship in the process is as follows:

[0075] ;

[0076] in, represents the upsampled feature map, Indicates the first smoke density map, An index representing the channel direction, Indicates the index in the high direction, Indicates the index in the width direction;

[0077] It should be noted that The process realizes the uniform grouping of the initial concentration map along the channel, and the 4×1×1 pixels in each group are used to randomly arrange into 1×2×2 spatial blocks, that is, compressing the channel to make up for the spatial information.

[0078] In the step of performing bilinear interpolation and upsampling operations on the smoke density maps of different levels except the first layer in sequence to obtain the upsampled density map, the corresponding relationship between the processes is as follows:

[0079] ;

[0080] in, represents the upsampled concentration map, Represents a bilinear interpolation operation;

[0081] In the step of performing a convolution operation on the upsampled feature map to obtain the convolution result of the upsampled feature map, performing a convolution operation and a global average pooling operation on the upsampled concentration map in sequence to obtain the average pooling result of the upsampled concentration map, and adding the convolution result of the upsampled feature map to the average pooling result to obtain the multi-scale feature group, the corresponding relationship between the process is as follows:

[0082] ;

[0083] in, represents a multi-scale feature group, represents a 1×1 convolution operation, Represents the point-by-point convolution operation of the third 1×1 window;

[0084] In the step of inputting the multi-scale feature group into the head module of the concentration map refinement module for splicing operation to obtain the smoke fine concentration map, the corresponding relationship of the process is as follows:

[0085] ;

[0086] in, Represents the smoke fine concentration map, represents the head network, Represents the first multi-scale feature group.

[0087] It should be noted that the present invention selects feature pyramid (FPN) as the head network. In practical applications, other commonly used segmentation head modules can be selected as the head network according to the task.

[0088] Step 5: reconstruct the smoke fine concentration map to obtain a reconstructed concentration map;

[0089] The reconstructed concentration map is fused with the smoke fine concentration map to obtain the final output credible concentration map.

[0090] See also Figure 5 In step 5, the smoke fine concentration map is reconstructed to obtain a reconstructed concentration map, which specifically includes the following steps:

[0091] The smoke fine concentration map is reconstructed using the self-encoding module to obtain a reconstructed map;

[0092] Combined with the reconstruction image construction error, the pixel-by-pixel reconstruction error is obtained, and the pixel-by-pixel reconstruction error is modeled to obtain the heteroscedastic generalized Gaussian distribution of the pixel-by-pixel reconstruction error. The parameters of the autoencoder module are expressed according to the heteroscedastic generalized Gaussian distribution of the pixel-by-pixel reconstruction error to obtain the likelihood expression of the autoencoder module parameters in the smoke fine concentration map and the smoke fine concentration map label;

[0093] Variance modeling is performed based on the heteroscedastic generalized Gaussian distribution of pixel-by-pixel reconstruction error to obtain scale parameters and shape parameters for uncertainty modeling;

[0094] Combining the scale parameter of uncertainty modeling, the shape parameter of uncertainty modeling and the pixel-by-pixel reconstruction error, the learning objectives of the reconstruction item of the Bayesian credible module and the learning objectives of the uncertainty item are obtained, and the autoencoder module parameters are gradient updated in the likelihood expression of the smoke fine concentration map and the smoke fine concentration map label to obtain the updated autoencoder module parameters;

[0095] The autoencoder module is updated using the updated parameters of the autoencoder module to obtain an updated autoencoder module;

[0096] Using the updated autoencoder module, the reconstructed concentration map is obtained.

[0097] In the step of reconstructing the smoke fine concentration map using the self-encoding module to obtain the reconstructed map, the corresponding relationship of the process is as follows:

[0098] ;

[0099] in, Represents the reconstructed graph, represents the autoencoder module, Represents the parameters of the autoencoder module;

[0100] In the step of combining the reconstruction map construction error to obtain the pixel-by-pixel reconstruction error, the pixel-by-pixel reconstruction error is modeled to obtain the heteroscedastic generalized Gaussian distribution of the pixel-by-pixel reconstruction error, the parameters of the autoencoder module are expressed according to the heteroscedastic generalized Gaussian distribution of the pixel-by-pixel reconstruction error, and the likelihood expression of the autoencoder module parameters in the smoke fine concentration map and the smoke fine concentration map label is obtained. The corresponding relationship between the process is as follows:

[0101] ;

[0102] in, represents all the labels in the dataset, represents the scale of the distribution, represents the shape of the distribution, represents the probability distribution, represents the number of samples, represents the heteroskedastic generalized Gaussian distribution, Indicates The labels of samples, Indicates The reconstruction of samples, Indicates The size of the sample, Indicates The shape of the samples;

[0103] It should be noted that the fine smoke concentration map calculated in steps 2 to 4 is approximated by the reconstructed concentration map.

[0104] In the step of performing variance modeling based on the heteroscedastic generalized Gaussian distribution of the pixel-by-pixel reconstruction error to obtain the scale parameter of uncertainty modeling and the shape parameter of uncertainty modeling, the corresponding relationship between the process is as follows:

[0105] ;

[0106] in, represents the variance, Indicates that it has been processed by the gamma function.

[0107] In the step of combining the scale parameter of uncertainty modeling, the shape parameter of uncertainty modeling and the pixel-by-pixel reconstruction error, the learning objectives of the reconstruction items of the Bayesian credible module and the learning objectives of the uncertainty items are obtained, and the autoencoder module parameters are used to perform gradient update on the likelihood expression of the smoke fine concentration map and the smoke fine concentration map label to obtain the updated parameters of the autoencoder module, the corresponding relationship of the process is as follows:

[0108] ;

[0109] in, represents the updated parameters of the autoencoder module, Indicates update parameters Minimize the formula, Represents logarithm.

[0110] It should be noted that is a set of optimal solutions that jointly minimize the reconstruction mean square error and negative log likelihood.

[0111] In the step of fusing the reconstructed concentration map with the smoke fine concentration map to obtain the final output credible concentration map, the corresponding relationship between the process is as follows:

[0112] ;

[0113] in, represents the credible concentration map of the final output, Represents the contribution coefficient.

[0114] It should be noted that In this embodiment, the contribution coefficient is 0.4.

[0115] For further information, see Figure 6 ,The test results are recorded in Tables 1 and 2, which respectively show the ,comparison results of the evaluation indicators of minimum mean square error and ,intersection over union ratio of segmentation.

[0116] Table 1: Minimum mean square error indicators of smoke concentration estimation by different methods

[0117]

[0118] Table 2: Intersection-over-Union ratio indicators of smoke pixel positioning and segmentation by different methods

[0119]

[0120] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0121] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0122] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A smoke concentration estimation method based on long- and short-range feature fusion, characterized in that: The method comprises the following steps: Step 1: construct a content extraction module based on depthwise separable convolution, and set up several layers of long and short distance feature modules in a stacked manner to construct a decoder; Step 2: Input the color smoke composite image into the content extraction module and encode it with the lightweight network to obtain the mid-level smoke features and the high-level smoke features respectively. Downsample the mid-level smoke features and concatenate them with the high-level smoke features to obtain the basic features of the smoke. Step 3: Input the basic features of the smoke into the information decoder, and use the long-distance and short-distance feature modules to decode the basic features of the smoke, and obtain the long-distance features and short-distance features of the smoke respectively; The long-distance features, short-distance features and basic features of smoke are fused to obtain the smoke density map of the current layer; The smoke density map of the current layer is used as the input of the long-short distance feature module of the next layer. After passing through several layers of long-short distance feature modules, smoke density maps of different levels are obtained. Step 4: perform spatial supplementation and upsampling operations on the channel information in the smoke density map of the first layer in sequence to obtain an upsampling feature map; Perform bilinear interpolation and upsampling operations on smoke density maps of different layers except the first layer in sequence to obtain upsampled density maps; Perform a convolution operation on the upsampled feature map to obtain the convolution result of the upsampled feature map, perform a convolution operation and a global average pooling operation on the upsampled concentration map in sequence to obtain the average pooling result of the upsampled concentration map, and add the convolution result of the upsampled feature map and the average pooling result to obtain a multi-scale feature group; The multi-scale feature group is input into the head module of the concentration map refinement module for splicing operation to obtain a fine smoke concentration map; Step 5: reconstruct the smoke fine concentration map to obtain a reconstructed concentration map; The reconstructed concentration map is fused with the smoke fine concentration map to obtain the final output credible concentration map.

2. The smoke density estimation method based on long- and short-range feature fusion according to claim 1 is characterized in that: In step 2, the color smoke composite image is input into the content extraction module and encoded in combination with the lightweight network to obtain the mid-level smoke features and the high-level smoke features, respectively. The mid-level smoke features are downsampled and concatenated with the high-level smoke features to obtain the basic features of the smoke. The corresponding relationship in the process is as follows: ； in, Indicates the basic characteristics of smoke, represents the splicing operation along the channel direction, represents the downsampling operation based on bilinear interpolation, Indicates the characteristics of mid-level smoke. Represents high-level smoke characteristics.

3. The smoke density estimation method based on long- and short-range feature fusion according to claim 2 is characterized in that: In step 3, the basic features of the smoke are input into the information decoder, and the basic features of the smoke are decoded using the long- and short-range feature modules to obtain the long-range features and short-range features of the smoke, respectively, which specifically includes the following sub-steps: Inputting the basic features of the smoke into the information decoder, performing average pooling operations and convolution operations in sequence, to obtain calculated components, wherein the calculated components include a first component, a second component, and a third component; The second component and the third component are subjected to matrix multiplication and normalization in sequence to obtain the extracted long-distance dependency; The extracted long-distance dependencies are subjected to random neuron discarding operations and matrix multiplication operations combined with the first component in sequence to obtain the long-distance features of the smoke; The basic features of smoke are input into the information decoder to perform group convolution operation, normalization operation along the channel direction and point-by-point convolution operation in sequence to obtain the intermediate local features; The intermediate local features are activated, convolved point by point, and randomly discarded neurons in sequence to obtain the short-range features of the smoke.

4. The smoke density estimation method based on long- and short-range feature fusion according to claim 3 is characterized in that: In the step of inputting the basic features of smoke into the information decoder and performing average pooling and convolution operations in sequence to obtain the calculated components, the corresponding relationship in the process is as follows: ； in, Indicates The calculated components, Indicates convolution operations, Indicates Average pooling operation; In the step of performing matrix multiplication and normalization on the second component and the third component in sequence to obtain the extracted long-distance dependency, the corresponding relationship in the process is as follows: ； in, represents the extracted long-distance dependency, represents normalization processing, represents a channel-by-channel multiplication operation, represents division, Indicates the square root of the number of channels that are dependent on long distances. represents the second component, represents the third component; In the step of performing random discard neuron operation and matrix multiplication operation combined with the first component on the extracted long-distance dependencies to obtain the long-distance features of the smoke, the corresponding relationship in the process is as follows: ； in, Indicates the long-distance characteristics of smoke, represents the random discarding of neuron operations, represents the first component; In the step of inputting the basic features of smoke into the information decoder and performing group convolution operation, normalization operation along the channel direction and point-by-point convolution operation in sequence to obtain the intermediate local features, the corresponding relationship in the process is as follows: ； in, represents the intermediate local features, represents the point-by-point convolution operation of the first 1×1 window, represents the normalization operation along the channel direction, Represents a grouped convolution operation of a 3×3 window; In the steps of performing activation operations, point-by-point convolution operations, and random neuron discarding operations on the intermediate local features in order to obtain the short-range features of smoke, the corresponding relationship between the processes is as follows: ； in, Indicates the short-range characteristics of smoke, represents the point-by-point convolution operation of the second 1×1 window, Represents a GeLU activation operation.

5. The smoke density estimation method based on long- and short-range feature fusion according to claim 4 is characterized in that: In the step of fusing the long-distance features, short-distance features and basic features of smoke to obtain the smoke density map of the current layer, the corresponding relationship between the process is as follows: ； in, Represents the smoke density map of the current layer. The fusion coefficient representing the long-distance characteristics of smoke, The fusion coefficient representing the short-range characteristics of smoke, The fusion coefficient representing the basic characteristics of smoke, represents the normalized coefficient, represents the indicator function, Indicates Coefficient .

6. The smoke density estimation method based on long- and short-range feature fusion according to claim 5 is characterized in that: In the step of using the smoke density map of the current layer as the input of the long-short distance feature module of the next layer, and obtaining smoke density maps of different levels after passing through several layers of long-short distance feature modules, the corresponding relationship of the process is as follows: ； in, Indicates Smoke density diagram of long and short distance feature modules, Indicates Smoke density diagram of long and short distance feature modules, Indicates Long and short distance feature modules.

7. The smoke density estimation method based on long- and short-range feature fusion according to claim 6 is characterized in that: In step 4, the channel information in the smoke density map of the first layer is sequentially subjected to spatial supplementation and upsampling operations to obtain an upsampling feature map. The corresponding relationship between the processes is as follows: ； in, represents the upsampled feature map, Represents the first layer of smoke density map, An index representing the channel direction, Indicates the index in the high direction, Indicates the index in the width direction; In the step of performing bilinear interpolation and upsampling operations on the smoke density maps of different levels except the first layer in sequence to obtain the upsampled density map, the corresponding relationship between the processes is as follows: ； in, represents the upsampled concentration map, Represents a bilinear interpolation operation; In the step of performing a convolution operation on the upsampled feature map to obtain the convolution result of the upsampled feature map, performing a convolution operation and a global average pooling operation on the upsampled concentration map in sequence to obtain the average pooling result of the upsampled concentration map, and adding the convolution result of the upsampled feature map to the average pooling result to obtain the multi-scale feature group, the corresponding relationship between the process is as follows: ； in, represents a multi-scale feature group, represents a 1×1 convolution operation, Represents the point-by-point convolution operation of the third 1×1 window; In the step of inputting the multi-scale feature group into the head module of the concentration map refinement module for splicing operation to obtain the smoke fine concentration map, the corresponding relationship of the process is as follows: ； in, Represents the smoke fine concentration map, represents the head network, Represents the first multi-scale feature group.

8. The smoke density estimation method based on long- and short-range feature fusion according to claim 7 is characterized in that: In step 5, the smoke fine concentration map is reconstructed to obtain a reconstructed concentration map, which specifically includes the following steps: The smoke fine concentration map is reconstructed using the self-encoding module to obtain a reconstructed map; Combined with the reconstruction image construction error, the pixel-by-pixel reconstruction error is obtained, and the pixel-by-pixel reconstruction error is modeled to obtain the heteroscedastic generalized Gaussian distribution of the pixel-by-pixel reconstruction error. The parameters of the autoencoder module are expressed according to the heteroscedastic generalized Gaussian distribution of the pixel-by-pixel reconstruction error to obtain the likelihood expression of the autoencoder module parameters in the smoke fine concentration map and the smoke fine concentration map label; Variance modeling is performed based on the heteroscedastic generalized Gaussian distribution of pixel-by-pixel reconstruction error to obtain scale parameters and shape parameters for uncertainty modeling; Combining the scale parameter of uncertainty modeling, the shape parameter of uncertainty modeling and the pixel-by-pixel reconstruction error, the learning objectives of the reconstruction item of the Bayesian credible module and the learning objectives of the uncertainty item are obtained, and the autoencoder module parameters are gradient updated in the likelihood expression of the smoke fine concentration map and the smoke fine concentration map label to obtain the updated autoencoder module parameters; The autoencoder module is updated using the updated parameters of the autoencoder module to obtain an updated autoencoder module; Using the updated autoencoder module, the reconstructed concentration map is obtained.

9. The smoke density estimation method based on long- and short-range feature fusion according to claim 8 is characterized in that: In the step of reconstructing the smoke fine concentration map using the self-encoding module to obtain the reconstructed map, the corresponding relationship of the process is as follows: ； in, Represents the reconstructed graph, represents the autoencoder module, Represents the parameters of the autoencoder module; In the step of combining the reconstruction map construction error to obtain the pixel-by-pixel reconstruction error, the pixel-by-pixel reconstruction error is modeled to obtain the heteroscedastic generalized Gaussian distribution of the pixel-by-pixel reconstruction error, the parameters of the autoencoder module are expressed according to the heteroscedastic generalized Gaussian distribution of the pixel-by-pixel reconstruction error, and the likelihood expression of the autoencoder module parameters in the smoke fine concentration map and the smoke fine concentration map label is obtained. The corresponding relationship between the process is as follows: ； in, represents all the labels in the dataset, represents the scale of the distribution, represents the shape of the distribution, represents the probability distribution, represents the number of samples, represents the heteroskedastic generalized Gaussian distribution, Indicates The labels of samples, Indicates The reconstruction of samples, Indicates The size of the sample, Indicates The shape of the samples; In the step of performing variance modeling based on the heteroscedastic generalized Gaussian distribution of the pixel-by-pixel reconstruction error to obtain the scale parameter of uncertainty modeling and the shape parameter of uncertainty modeling, the corresponding relationship between the process is as follows: ； in, represents the variance, Indicates that it has been processed by the gamma function; In the step of combining the scale parameter of uncertainty modeling, the shape parameter of uncertainty modeling and the pixel-by-pixel reconstruction error, the learning objectives of the reconstruction items of the Bayesian credible module and the learning objectives of the uncertainty items are obtained, and the autoencoder module parameters are used to perform gradient update on the likelihood expression of the smoke fine concentration map and the smoke fine concentration map label to obtain the updated parameters of the autoencoder module, the corresponding relationship of the process is as follows: ； in, represents the parameters of the updated autoencoder module, Indicates update parameters Minimize the formula, Represents logarithm.

10. The smoke density estimation method based on long- and short-range feature fusion according to claim 9, characterized in that: In the step of fusing the reconstructed concentration map with the smoke fine concentration map to obtain the final output credible concentration map, the corresponding relationship between the process is as follows: ； in, represents the credible concentration map of the final output, Represents the contribution coefficient.

Images