A dynamic smoke screen image recognition method for a smoke generator

Abstract

This invention provides a method for recognizing dynamic smoke screen images from smoke generators. It utilizes the DPAUNet model to extract and optimize deep features from dynamic smoke screen images, and combines this with the LogSAD algorithm to accurately measure feature similarity, thus constructing a collaborative optimization mechanism for feature extraction and similarity measurement. The invention includes the following steps: Step 1: Acquire dynamic smoke screen images during the operation of the smoke generator, and preprocess the acquired images by converting the acquired color images to grayscale images; Step 2: Utilize the DPAUNet model, combined with deformable convolution and a parameter-free attention mechanism, to extract and optimize the deep latent features of the preprocessed dynamic smoke screen images, removing redundant features; Step 3: Introduce weighting factors and logarithmic transformation into the LogSAD algorithm to calculate the feature similarity between the image to be recognized and the smoke screen template image. Compare this feature similarity with a recognition threshold to determine whether the image to be recognized is a dynamic smoke screen image generated by the smoke generator.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically to a method for recognizing dynamic smoke screen images of smoke generators. Background Technology

[0002] As a specialized device capable of rapidly generating large-scale smoke screens, the coverage area, concentration distribution, diffusion speed, and duration of the dynamic smoke screen produced by a smoke generator directly determine the effectiveness evaluation of its application scenarios. Therefore, real-time and accurate image recognition of the dynamic smoke screen generated during the operation of the smoke generator, and the acquisition of its dynamic change characteristics, is a core prerequisite for quantitative evaluation of the smoke generator's application effect, monitoring of its operational status, and optimized control, possessing extremely high practical application value. Currently, dynamic smoke screen image recognition technology has become a research hotspot in related fields, but existing technologies still have many shortcomings and are insufficient to meet the actual needs of dynamic smoke screen recognition in smoke generators.

[0003] First, existing dynamic smokescreen recognition methods suffer from insufficient feature extraction capabilities and poor anti-interference performance. Dynamic smokescreens are inherently characterized by irregular shapes, uneven concentration distribution, and complex dynamic diffusion characteristics. Furthermore, real-world applications often encounter interference factors such as changing lighting, background occlusion, and environmental noise. Existing methods often employ a single feature extraction approach, neglecting the spatiotemporal correlation features of the smokescreen during dynamic diffusion and failing to effectively distinguish smokescreen features from background interference features. This leads to feature redundancy and loss of key features in complex scenarios, thus affecting recognition accuracy. Second, existing methods lack a collaborative optimization mechanism for feature extraction and similarity measurement. The effectiveness of feature extraction and the suitability of similarity measurement directly determine the overall performance of the recognition method. Existing methods often design these two as independent steps without collaborative optimization, resulting in extracted features that are not adequately adapted to the subsequent similarity measurement process, thus impacting recognition performance.

[0004] Furthermore, existing dynamic smoke recognition methods lack generalization ability and struggle to adapt to the smoke recognition needs of different types of smoke generators and application scenarios. Different smoke generators exhibit variations in smoke generation rate, smoke concentration, and smoke particle size, while the lighting and background environments of different application scenarios also differ. Existing methods are mostly designed for specific scenarios and types of smoke generators, lacking adaptive adjustment capabilities. When the scenario or type of smoke generator changes, the recognition performance drops significantly, resulting in poor versatility and hindering widespread engineering application. Therefore, there is an urgent need for a dynamic smoke image recognition method specifically designed for smoke generators, adaptable to the dynamic smoke recognition needs of different application scenarios and types of smoke generators. Summary of the Invention

[0005] To address the shortcomings of existing dynamic smoke screen image recognition methods, such as insufficient feature extraction capabilities, poor adaptability of similarity measurement, insufficient collaborative optimization, and weak generalization ability, this invention provides a dynamic smoke screen image recognition method for smoke generators. It utilizes the DPAUNet model to extract and optimize deep features of dynamic smoke screen images, combines the LogSAD algorithm to achieve accurate measurement of feature similarity, and constructs a collaborative optimization mechanism for feature extraction and similarity measurement. This improves the robustness, accuracy, and versatility of dynamic smoke screen recognition, adapting to the dynamic smoke screen recognition needs of different application scenarios and types of smoke generators, and providing technical support for evaluating the application effect and monitoring the working status of smoke generators.

[0006] The objective of this invention can be achieved through the following technical solution: a dynamic smoke screen image recognition method for smoke generators, comprising the following steps: Step 1: Collect dynamic smoke screen images during the operation of the smoke generator, preprocess the collected images, and convert the collected color images into grayscale images; Step 2: Using the DPAUNet model, combined with deformable convolution and parameter-free attention mechanisms, extract deep latent features from the preprocessed dynamic smoke image and filter and optimize them, removing redundant features; including the following steps: Step 200: Construct the DPAUNet model; Step 201: Input the grayscale image preprocessed in Step 1 into the input layer of the DPAUNet model. The image is processed by the encoder bottleneck block in sequence to gradually extract the shallow detail features and deep semantic features of the dynamic smoke screen. Step 202: After the bottleneck block of each encoder layer is processed, the importance score of each pixel in the feature map is calculated by the SimAM parameterless attention module. Key smoke features and background interference features are selected to obtain the optimized feature map of each layer. The optimized feature maps of each layer of the encoder are passed to the corresponding layer of the decoder through skip connections. Step 3: Introduce a weighting factor and logarithmic transformation into the LogSAD algorithm to calculate the feature similarity between the image to be identified and the smoke template image. Compare this feature similarity with the recognition threshold to determine whether the image to be identified is a dynamic smoke screen image generated by the smoke generator.

[0007] Step 1 of the present invention includes the following steps: Step 100: Acquire dynamic smoke screen images during the operation of the smoke generator using an image acquisition device. The acquired images are color RGB images, and the image sequence is stored in a standard image format. Step 101: Denoise the acquired color image using Gaussian filtering. Step 102: Convert the denoised color image to a grayscale image; Step 103: Normalize the grayscale image and map it to the [0, 1] interval.

[0008] Step 101 of the present invention includes the following steps: First, adjust the standard deviation σ of the Gaussian filter based on the noise intensity of the acquired image; Then, determine the size of the Gaussian filter convolution kernel; Finally, the convolution kernel is convolved with each pixel of the acquired color image, and the pixel value G(x,y) of the denoised dynamic smoke image at coordinates (x,y) is calculated according to the Gaussian filtering formula: , Where (u,v) are the center coordinates of the Gaussian filter convolution kernel, and f(x,y) is the pixel value of the original smoke image at coordinates (x,y) before denoising; In step 102, a weighted average method is used for grayscale processing. The calculation formula for grayscale processing is as follows: Gray(x,y)=R1×R(x,y)+G1×G(x,y)+B1×B(x,y), Where Gray(x,y) is the gray value of the image at coordinates (x,y) after grayscale conversion, R(x,y), G(x,y), and B(x,y) are the pixel values ​​of the red, green, and blue channels of the denoised color image at coordinates (x,y), respectively, and R1, G1, and B1 are the weight coefficients of the red, green, and blue channels, respectively. Step 103 includes the following steps: Traverse the grayscale image and find the minimum pixel value G in the grayscale image. min and the maximum pixel value G max Then, according to the normalization formula, each pixel value Gray(x,y) is converted into a normalized pixel value G. norm (x, y), the calculation formula is as follows: .

[0009] The bottleneck block feature calculation formula of the DPAUNet model described in this invention is as follows: (1) The formula for 3×3 convolution feature compression is: F 3×3 =Conv 3×3 (X) in W 3×3 ,b 3×3 ), In the formula, X in W is the input feature map for the bottleneck block. 3×3 b 3×3 The weights and biases of the 3×3 convolution are respectively, F 3×3The compressed feature map output by the 3×3 convolution, Conv 3×3 This represents a standard convolution operation with a 3×3 dimension. (2) The adaptive capture formula for deformable convolution V2 features is: F dconv =DConv V2 (F) 3×3 W d ,Δp,Δm). In the formula, W d F represents the convolution weights of deformable convolution V2, Δp is the adaptive offset, Δm is the offset weight, and F... dconv DConv outputs feature maps for deformable convolution V2. V2 This represents the deformable convolution V2 operation; (3) The formula for recovering 1×1 convolutional features is: F 1×1 =Conv 1×1 (F) dconv W 1×1 ,b 1×1 ), In the formula, W 1×1 b 1×1 The weights and biases of the 1×1 convolution are respectively, F 1×1 Conv is the feature map of the recovered channel at the final output of the bottleneck block. 1×1 This represents a standard convolution operation of size 1×1.

[0010] In step 202 of this invention, the importance scoring formula for the SimAM parameterless attention module is: , In the formula, s i,j c represents the importance score of the j-th pixel in the i-th layer feature map; i,j Let μ be the pixel value of the j-th feature pixel in the i-th layer feature map. i Let v be the average pixel value of the feature map of the i-th layer. i Let σ be the global extremum of the pixel in the i-th layer feature map. i Let be the standard deviation of the pixels in the feature map of the i-th layer. To prevent the minimum value where the denominator is 0.

[0011] In step 202 of this invention, the formula for the feature map after screening by the SimAM parameterless attention module is: F opt,i =Softmax(S i )·F enc,i , In the formula, F opt,iS is the optimized feature map of the i-th layer encoder after being filtered by the SimAM parameterless attention module. i The SimAM attention importance score for the feature map of layer i is determined by the sum of the values ​​of all pixels in that layer. i,j It is assembled according to position, and Softmax is a normalization operation. enc,i For the original feature map output by the i-th layer encoder, Softmax(S) i ) is for S i Perform normalization.

[0012] In step 200 of this invention, the DPAUNet model is a lightweight variant of UNet, and its core consists of an encoder, a decoder, deformable convolution V2, a SimAM parameterless attention module, and a skip connection structure. The DPAUNet model construction process is as follows: The input to the DPAUNet model is a preprocessed grayscale image; the encoder adopts a 4-layer lightweight bottleneck block structure, each bottleneck block consisting of a 3×3 convolutional layer, a deformable convolutional V2 layer, and a 1×1 convolutional layer; each bottleneck block is followed by a BatchNorm layer and a ReLU activation function. After the SimAM parameterless attention module is embedded in the bottleneck block of each layer of the encoder, no additional training parameters are required. By calculating the importance score of each pixel in the feature map, key smoke features are selected and background interference features are suppressed. The attention score is calculated based on the local contrast and global distribution of feature pixels. The decoder adopts a 4-layer deconvolution structure symmetrical to the encoder. At the same time, it fuses the shallow detail features extracted by the corresponding layer of the encoder with the deep semantic features of the decoder through skip connections. The output of the DPAUNet model is an optimized deep feature vector. The dimension of the feature vector is adaptively set according to the input image size and encoder structure. The output feature vector has completed the filtering of redundant features, retaining only the key features that can effectively represent the dynamic smoke screen.

[0013] The training process of the DPAUNet model described in this invention is as follows: Multiple sets of dynamic smoke screen images and non-smoke screen images from smoke generators are collected as training samples. The dynamic smoke screen images cover different types of smoke generators, different smoke concentrations, different diffusion stages, and different application scenarios. The non-smoke screen images cover background images similar to the application scenarios of the smoke generators, serving as negative samples. All training samples are preprocessed sequentially using step 1 to obtain standardized training sample images. Using the preprocessed training sample images as input and the sample labels as supervision signals, the DPAUNet model is trained using the cross-entropy loss function and the Adam optimizer. Iterative training continues until the model loss value converges and the recognition accuracy reaches a preset threshold, thus completing the training of the DPAUNet model.

[0014] Step 3 of this invention includes the following steps: Step 300: Pre-collect multiple sets of dynamic smoke screen images during normal operation of the smoke generator as template samples; perform the preprocessing in Step 1 and the DPAUNet model feature extraction and optimization in Step 2 on each set of template samples to obtain the optimized feature vector of each set of template samples. Take the average of the optimized feature vectors of all template samples to obtain the baseline feature vector F of the smoke screen template. temp The baseline eigenvector F temp The dimension of the image to be recognized and the optimized feature vector F opt The dimensions are consistent; Step 301: Based on the importance score output by the SimAM parameterless attention module in step 202, calculate the optimized feature vector F of the image to be recognized. opt The weight factor w for each feature i ; Step 302, F opt F temp and w i Substitute the values ​​into the LogSAD algorithm formula to calculate the similarity value ILogSAD between the image to be identified and the template image; Step 303: Calculate the average similarity value between multiple sets of template feature vectors and use it as the recognition threshold T; Step 304: Compare the similarity value ILogSAD with the recognition threshold T to determine whether the image to be identified is a dynamic smoke screen image generated by a smoke generator. The judgment rules are as follows: If ILogSAD ≤ T, then the image to be identified is a dynamic smoke screen image generated by a smoke generator; If ILogSAD>T, then the image to be identified is determined to be a non-smoke image or an abnormal smoke image.

[0015] The formula for calculating the similarity value ILogSAD as described in this invention is: , In the formula, F opt,i F is the optimized feature vector of the image to be recognized. opt The i-th element, F temp,i It is the feature vector F after optimizing the smoke screen template image. temp The i-th element.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention is based on the LogSAD algorithm, and introduces logarithmic transformation and weighting factors. The logarithmic transformation enhances the sensitivity to smoke screen details and small errors, reduces the impact of large errors, and avoids measurement deviation caused by blurring of smoke screen edges. The weighting factors are allocated according to the importance of features, highlighting the differences in key smoke screen features, reducing the impact of background interference and noise, solving the problems of poor adaptability, low measurement accuracy and weak anti-interference ability of traditional SAD methods, and improving the accuracy and robustness of feature similarity measurement.

[0017] (2) This invention constructs a collaborative optimization mechanism for DPAUNet feature extraction optimization and LogSAD similarity measurement. DPAUNet adapts to the spatial irregularity of dynamic smoke screens through deformable convolution V2 and filters key features through SimAM parameterless attention. The extracted optimized features can fully adapt to the measurement requirements of LogSAD. The design of LogSAD weight factors is related to the feature importance score of DPAUNet, realizing the collaborative improvement of feature extraction and similarity measurement, which greatly improves the overall performance of dynamic smoke screen recognition and can adapt to complex scenes and complex changes in dynamic smoke screens. At the same time, the lightweight nature of DPAUNet reduces the computational complexity and adapts to the engineering real-time recognition requirements. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an embodiment of the present invention.

[0019] Figure 2 This is a flowchart illustrating step 3 of an embodiment of the present invention.

[0020] Figure 3 This is a training loss curve in an experiment according to an embodiment of the present invention.

[0021] Figure 4 This is a graph showing the accuracy of the Dice similarity coefficient on the validation set in the experiment of an embodiment of the present invention.

[0022] Figure 5 This is a graph showing the accuracy of the average crossover-union ratio on the validation set in the experiment of an embodiment of the present invention. Detailed Implementation

[0023] To facilitate understanding of the embodiments of this application, the following will provide further explanation and description with reference to the accompanying drawings and specific embodiments. These embodiments do not constitute a limitation on the embodiments of this application.

[0024] I. Implementation Examples

[0025] See Figure 1 This embodiment provides a dynamic smoke screen image recognition method for smoke generators, which includes the following steps: Step 1: Acquire dynamic smoke screen images during the operation of the smoke generator. Preprocess the acquired images by converting the color images to grayscale images. The preprocessing steps include filtering and noise reduction, grayscale conversion, and normalization, providing standardized support for subsequent feature extraction. Specifically, this includes the following steps: Step 100: Select an industrial camera as the image acquisition device and install it in a suitable position around the smoke generator to ensure that the dynamic smoke diffusion process during the operation of the smoke generator can be fully captured. The industrial camera acquires dynamic smoke images during the operation of the smoke generator. The acquisition frame rate is adaptively adjusted according to the smoke generation rate of the smoke generator to ensure that the dynamic changes of the smoke are captured. The acquired images are color RGB images, and the image resolution is adaptively set according to the actual monitoring requirements. The acquired image sequence is stored in a standard image format for subsequent preprocessing.

[0026] Step 101: Gaussian noise and random noise are inevitably present in the acquired color dynamic smoke image. These noises will interfere with the accuracy of subsequent feature extraction. Therefore, Gaussian filtering is used to denoise the acquired color image, including the following steps: First, based on the noise intensity of the acquired image, the standard deviation σ of the Gaussian filter is adaptively adjusted. The greater the noise intensity, the larger the value of σ, ensuring effective noise removal. Next, the size of the Gaussian filter convolution kernel is determined, adaptively matched to the image resolution to ensure that key details of the smoke screen are not lost during the filtering process. Finally, the convolution kernel is convolved with each pixel of the acquired color image, and the pixel value G(x,y) of the denoised smoke screen image at coordinates (x,y) is calculated according to the Gaussian filter formula, completing the image denoising. The denoised image retains the key features of the smoke screen while eliminating noise interference. The pixel value G(x,y) is calculated as follows: In the formula, (u,v) are the center coordinates of the Gaussian filter convolution kernel, and f(x,y) is the pixel value of the original smoke image at coordinates (x,y) before denoising.

[0027] Step 102: Convert the denoised color image to a grayscale image. A weighted average method is used for grayscale conversion, preserving the key grayscale features of the smoke screen while reducing the dimensionality of the image data and decreasing computational load. The grayscale conversion formula is as follows: Gray(x,y) = R1×R(x,y) + G1×G(x,y) + B1×B(x,y), where Gray(x,y) is the gray value of the image at coordinates (x,y) after grayscale conversion, and R(x,y), G(x,y), and B(x,y) are the pixel values ​​of the red, green, and blue channels of the denoised color image at coordinates (x,y), respectively. R1, G1, and B1 are the weighting coefficients for the red, green, and blue channels, respectively.

[0028] Step 103: Normalize the grayscale image and map it to the [0, 1] interval.

[0029] Traverse the grayscale image and find the minimum pixel value G in the grayscale image. min and the maximum pixel value G max Then, according to the normalization formula, each pixel value Gray(x,y) is converted into a normalized pixel value G. norm (x,y) eliminates the influence of scale differences, ensuring the accuracy of subsequent feature extraction. The normalization formula is as follows: .

[0030] Step 2: Using the DPAUNet model, combined with deformable convolution and parameter-free attention mechanisms, extract deep latent features from the preprocessed dynamic smoke image and perform filtering and optimization to remove redundant features. Specifically, this includes the following steps: Step 200: Construct the DPAUNet model.

[0031] The DPAUNet model is a lightweight variant of UNet, with its core consisting of an encoder, decoder, deformable convolution V2, a SimAM parameterless attention module, and skip connection structures. It is used for deep feature extraction and redundant feature filtering of dynamic smoke screen images. The specific model construction process is as follows: The DPAUNet model takes a preprocessed, normalized grayscale image as input. The image size is adaptively adjusted according to the resolution of the actual acquired smoke screen image to ensure that the model input matches the image size. The encoder adopts a 4-layer lightweight bottleneck block structure. Each bottleneck block consists of a 3×3 convolutional layer, a deformable convolutional V2 layer, and a 1×1 convolutional layer. The 3×3 convolutional layer is used to compress the number of feature channels and reduce computation. The deformable convolutional V2 layer is used to capture the spatial adaptive features of dynamic smoke screens, adapting to the irregular shape and dynamic diffusion characteristics of smoke screens. The 1×1 convolutional layer is used to restore the number of feature channels, achieving feature dimensionality reduction while retaining key features. Each bottleneck block is followed by a BatchNorm layer and a ReLU activation function to introduce nonlinear transformation, enhance the model's feature learning ability, and suppress overfitting.

[0032] The core formula for calculating the bottleneck block features in DPAUNet is as follows: (1) The formula for 3×3 convolution feature compression is: F 3×3 =Conv 3×3 (X) in W 3×3 ,b 3×3 In the formula, X in W is the input feature map for the bottleneck block.3×3 b 3×3 The weights and biases of the 3×3 convolution are respectively, F 3×3 The compressed feature map output by the 3×3 convolution, Conv 3×3 This represents a standard convolution operation with a 3×3 size.

[0033] (2) The adaptive capture formula for deformable convolution V2 features is: F dconv =DConv V2 (F) 3×3 W d ,Δp,Δm), where W d F represents the convolution weights of deformable convolution V2, Δp is the adaptive offset, Δm is the offset weight, and F... dconv DConv outputs feature maps for deformable convolution V2. V2 This indicates the deformable convolution V2 operation.

[0034] (3) The formula for recovering 1×1 convolutional features is: F 1×1 =Conv 1×1 (F) dconv W 1×1 ,b 1×1 In the formula, W 1×1 b 1×1 The weights and biases of the 1×1 convolution are respectively, F 1×1 Conv is the feature map of the recovered channel at the final output of the bottleneck block. 1×1 This represents a standard convolution operation of size 1×1.

[0035] After the SimAM parameterless attention module is embedded in the bottleneck block of each layer of the encoder, no additional training parameters are required. It selects key smoke screen features and suppresses background interference features by calculating the importance score of each pixel in the feature map. The attention score calculation is based on the local contrast and global distribution of feature pixels to ensure the rationality and relevance of the score. The decoder adopts a 4-layer deconvolution structure symmetrical to the encoder to upsample the deep features extracted by the encoder, restore the spatial dimension of the image features, and fuse the shallow detail features extracted by the corresponding layer of the encoder with the deep semantic features of the decoder through skip connections to make up for the feature loss during the upsampling process and optimize the feature expression effect.

[0036] The importance scoring formula for the SimAM parameterless attention module is as follows: In the formula, s i,j c represents the importance score of the j-th pixel in the i-th layer feature map; i,j Let μ be the pixel value of the j-th feature pixel in the i-th layer feature map. iLet v be the average pixel value of the feature map of the i-th layer. i Let σ be the global extremum of the pixel in the i-th layer feature map. i Let be the standard deviation of the pixels in the i-th layer feature map. To prevent the minimum value where the denominator is 0.

[0037] The output of the DPAUNet model is an optimized deep feature vector. The dimension of the feature vector is adaptively set according to the input image size and encoder structure. The output feature vector has completed the filtering of redundant features, retaining only the key features that can effectively represent the dynamic smoke screen, which is suitable for the similarity measurement requirements of the subsequent LogSAD algorithm.

[0038] Step 201: Input the normalized grayscale image after preprocessing in Step 1 into the input layer of the DPAUNet model. It is processed by four encoder bottleneck blocks in sequence. Each bottleneck block uses a 3×3 convolution to compress the channel, deformable convolution V2 to capture spatial features, and a 1×1 convolution to restore the channel. Combined with the nonlinear transformation of the BatchNorm layer and the ReLU activation function, the shallow detail features and deep semantic features of the dynamic smoke screen are gradually extracted.

[0039] The core formula for multi-layer feature extraction of the encoder is: F enc,i =ReLU(BatchNorm(F 1×1,i )), In the formula, F enc,i F is the feature map output by the i-th layer encoder. 1×1,i This is the feature map after recovering the channels of the 1×1 convolution output of the bottleneck block of the i-th encoder. BatchNorm is a batch normalization operation used to suppress overfitting, and ReLU is an activation function used to introduce non-linear transformations to enhance feature learning ability. ReLU(BatchNorm(F)) 1×1,i )) is the feature map F after recovering the channels of the 1×1 convolution in the bottleneck block of the i-th encoder layer. 1×1,i First, batch normalization is performed, then a nonlinear transformation is introduced through the ReLU activation function, finally obtaining the feature map F output by the i-th layer encoder. enc,i .

[0040] Step 202: After the bottleneck block of each encoder layer is processed, the importance score of each pixel in the feature map is calculated by the SimAM parameterless attention module. Key smoke features and background interference features are selected to obtain the optimized feature map of each layer. The optimized feature maps of each layer of the encoder are passed to the corresponding layer of the decoder through skip connections to prepare for subsequent feature fusion.

[0041] The formula for the feature map after filtering by the SimAM parameterless attention module is: F opt,i =Softmax(Si )·F enc,i , In the formula, F opt,i S is the optimized feature map of the i-th layer encoder after being filtered by the SimAM parameterless attention module. i The SimAM attention importance score for the feature map of layer i is determined by the sum of the values ​​of all pixels in that layer. i,j It is constructed by concatenating positions, and Softmax is a normalization operation used to map the attention score to the [0,1] interval. enc,i This represents the original feature map output by the i-th layer encoder. Softmax(S) i S is the SimAM attention importance score of the feature map of layer i. i Perform normalization.

[0042] Step 203: The decoder upsamples the deepest feature map output by the encoder through a deconvolution layer, gradually restoring the spatial dimension of the features. Simultaneously, it fuses the optimized feature maps from the corresponding layers passed through skip connections with the feature map of the current layer in the decoder, compensating for feature loss during the upsampling process. Decoder feature fusion formula: F dec,i =DeConv(F dec,i+1 )⊕F opt,4-i+1 .

[0043] In the formula, F dec,i Let F be the fused feature map output from the i-th layer decoder, DeConv be the deconvolution operation, and F be the deconvolution operation. dec,i+1 Let F be the output feature map of the (i+1)th layer decoder, ⊕ be the feature fusion operation, and F be the output feature map of the (i+1)th layer decoder. opt,4-i+1 This is the optimized feature map passed to the corresponding layer of the encoder. DeConv(F) dec,i+1 ) is the feature map F output by the (i+1)th layer decoder. dec,i+1 Perform deconvolution upsampling. Since the encoder of this invention has a total of 4 layers, and the encoder and decoder are designed as a symmetrical architecture, the encoder layers are ordered from shallow to deep in a forward direction, and the decoder layers are arranged in a forward direction according to the upsampling iteration order. To ensure that the i-th layer of the decoder accurately matches the symmetrical feature layer corresponding to the encoder resolution and spatial size, a reverse index mapping of the encoder layers is required. The constant 4 in the formula represents the total number of encoder layers. Through the reverse calculation method of 4-i+1, the encoder optimized feature layer F corresponding one-to-one with the current layer of the decoder can be located in reverse. opt,4-i+1 This ensures that the feature maps of skip connections are of uniform scale and dimensionally adapted, enabling effective complementarity and fusion of shallow detail features and deep semantic features.

[0044] After processing by the 4-layer decoder, the final feature map is converted into a one-dimensional feature vector through a global average pooling layer, which is the optimized deep feature vector F output by the DPAUNet model.opt The similarity metric used in the subsequent LogSAD algorithm: F opt =GAP(F dec,4 ).

[0045] In the formula, F opt The optimized deep feature vector is the final output of the DPAUNet model. GAP is a global average pooling operation used to convert the final two-dimensional feature map F from the decoder. dec,4 Convert to a one-dimensional feature vector to suit subsequent similarity measurements. GAP(F) dec,4 ) is the two-dimensional fused feature map F output from the 4th layer of the decoder. dec,4 Perform a global average pooling operation.

[0046] Step 204: Training process of the DPAUNet model: Collect multiple sets of dynamic smoke screen images and non-smoke screen images from smoke generators as training samples. The dynamic smoke screen images cover different types of smoke generators, different smoke concentrations, different diffusion stages, and different application scenarios to ensure the representativeness and comprehensiveness of the samples. The non-smoke screen images cover background images similar to the application scenarios of smoke generators and serve as negative samples. Perform the preprocessing of all training samples in step 1 to obtain standardized training sample images. Use the preprocessed training sample images as input and the sample labels (smoke screen / non-smoke screen) as supervision signals. Use the cross-entropy loss function and Adam optimizer to train the DPAUNet model. During the training process, adaptively adjust the learning rate to avoid gradient vanishing or gradient exploding. Iterate the training until the model loss value converges and the recognition accuracy reaches the preset threshold to complete the training of the DPAUNet model. The trained model can be directly used for feature extraction and optimization of dynamic smoke screen images. The trained DPAUNet model is used to perform dynamic smoke image feature extraction, optimization, and one-dimensional feature vector transformation operations in steps 201-203, providing adaptive feature input for subsequent LogSAD similarity measurement.

[0047] Step 3: Introduce a weighting factor and logarithmic transformation into the LogSAD algorithm to calculate the feature similarity between the image to be identified and the smoke template image. Compare this feature similarity with a recognition threshold to determine whether the image to be identified is a dynamic smoke screen image generated by a smoke generator. See also... Figure 2 Specifically, it includes the following steps: Step 300: Pre-collect multiple sets of dynamic smoke screen images during normal operation of the smoke generator as template samples. The template samples cover smoke screen images with different smoke concentrations and different diffusion stages to ensure the representativeness of the template samples. Perform the preprocessing of Step 1 and the DPAUNet feature extraction and optimization of Step 2 on each set of template samples to obtain the optimized feature vector of each set of smoke screen template samples. Take the average of the optimized feature vectors of all template samples to obtain the baseline feature vector F of the smoke screen template. temp F temp The dimension of the image to be recognized and the optimized feature vector F opt The dimensions are consistent, and the smoke screen template feature vector F temp A benchmark used for subsequent similarity measurements.

[0048] Step 301: Based on the importance score output by the SimAM parameterless attention module in step 202, calculate the optimized feature vector F of the image to be recognized. opt The weight factor w for each feature i This includes the following steps: First, extract the optimized feature vector F. opt The feature importance score S for each feature in the dataset. i Then, calculate the sum of all feature importance scores. Where L3 is the optimized feature vector F opt and template feature vector F temp The dimensions of both are consistent. Finally, the weight factor w for each feature is calculated based on the weight factor. i Weighting factor w i The value range is [0,1], and the sum of all weight factors is 1, ensuring that key features have higher weight in the similarity measure and highlighting the differences in key features. Weight factor w i The calculation method is as follows: .

[0049] Step 302: Calculate the optimized feature vector F of the image to be recognized. opt Smoke template feature vector F temp and the calculated weighting factor w i Substituting into the LogSAD algorithm formula, the similarity value ILogSAD between the image to be recognized and the template image is calculated. The calculation formula is as follows: .

[0050] Among them, F opt,i F is the optimized feature vector of the image to be recognized. opt The i-th element, F temp,i It is the feature vector F after optimizing the smoke screen template image. tempThe i-th element. During the calculation, the logarithmic transformation ln(1+|F... opt,i -F temp,i |) This enhances sensitivity to small errors. When the feature differences between the image to be identified and the template image are small, the value after logarithmic transformation changes more significantly, effectively distinguishing subtle differences in the smoke screen. When the feature differences are large, the value after logarithmic transformation increases more gradually, avoiding distortion of the similarity value due to excessive differences in individual features; weighting factor w i The introduction of this feature makes the differences in key features dominate the similarity calculation, reduces the influence of background interference features and invalid features, and improves the accuracy and robustness of similarity measurement.

[0051] Step 303: Calculate the average similarity value between multiple sets of template feature vectors and use it as the recognition threshold T. The recognition threshold T is adaptively adjusted according to the actual application scenario and the type of smoke generator. There is no need to set a specific value to ensure the adaptability of the recognition threshold.

[0052] Step 304: Compare the similarity value ILogSAD between the image to be identified and the template image calculated in the above steps with the recognition threshold T to determine whether the image to be identified is a dynamic smoke screen image generated by the smoke generator. The judgment rules are as follows: If ILogSAD ≤ T, then the image to be identified is a dynamic smoke screen image generated by a smoke generator; If ILogSAD>T, then the image to be identified is determined to be a non-smoke image or an abnormal smoke image.

[0053] By outputting the recognition and judgment results, the method's generalization ability is ensured, adapting to the dynamic smoke screen recognition needs of different scenarios and different types of smoke generators.

[0054] II. Experiment.

[0055] 1. Training results.

[0056] The dynamic smoke diffusion process of the smoke generator was recorded in real time using an image acquisition device. A sampling frequency of 10 frames per second was set to extract keyframe sequences from the video stream, ultimately obtaining 1392 dynamic smoke images to construct a sample dataset. For the acquired dynamic smoke images, a semi-supervised annotation strategy combining automated pre-annotation scripts with manual fine-tuning was employed to complete the category labels (smoke / non-smoke) and key feature region annotations for the smoke images. Furthermore, the constructed annotation dataset was randomly divided into training, validation, and test sets in a 7:1:2 ratio to support the effective training and generalization performance evaluation of the method of this invention.

[0057] To ensure the efficiency of training and the reproducibility of experimental results, the hardware platform is equipped with a high-performance graphics processor (NVIDIA GeForce RTX 4090 24GB VRAM) to meet the large-scale tensor computation requirements of the method in pixel-level feature extraction and multi-scale feature fusion. It also features a high-frequency multi-core CPU and 64GB of RAM to ensure rapid image data reading and preprocessing. Based on the Ubuntu 20.04 operating system, the network was built and its parameters optimized using the Python 3.8 programming language and the PyTorch deep learning framework. CUDA and the accompanying cuDNN deep neural network acceleration library were used to accelerate the GPU's underlying computing power. Based on the above hardware and software environment, a total of 20 epochs of iterative training were conducted. To intuitively verify the convergence of the training process and the effectiveness of network optimization, training loss curves were plotted as follows: Figure 3 As shown. By Figure 3 As can be seen, this invention exhibits efficient feature mapping and learning capabilities in the early stages of training, with the training loss value dropping sharply from around 2.3 initially to around 0.5 within the first five rounds. Subsequently, this invention enters a stable parameter optimization phase for deep features, with the loss curve showing a continuous and gradual downward trend, smoothly converging and stabilizing at around 0.1 by the 20th round. This indicates that the network structure design of the method in this invention is reasonable, effectively fitting the complex spatial distribution characteristics of dynamic smoke screens, and without exhibiting training anomalies such as gradient divergence.

[0058] To comprehensively and objectively quantify and evaluate the performance of this invention in pixel-level recognition tasks, this study introduces the Dice Similarity Coefficient (Dice Score) and Mean Intersection over Union (mIoU) as core evaluation metrics. The Dice Score primarily measures the spatial overlap between the prediction mask and the ground truth mask, exhibiting high sensitivity to local boundary details of the target region, particularly demonstrating significant robustness in scenarios with severe imbalances in foreground and background pixel ratios. mIoU, on the other hand, calculates the ratio of the intersection to the union of the predicted and ground truth regions for each category, and averages the results globally, rigorously characterizing the overall accuracy of this invention in pixel-level classification. Both metrics range from 0 to 1; values ​​closer to 1 indicate richer and more accurate multi-scale features extracted by this invention, resulting in superior region segmentation accuracy and generalization performance. The Dice Similarity Coefficient and Mean Intersection over Union results on the validation set are shown below. Figure 4 and Figure 5 As shown.

[0059] Depend on Figure 4 It can be seen that the Dice score exhibits a positive correlation with the decrease in training loss. In the first 10 training rounds, as shallow details and deep semantic features are gradually integrated, the Dice score of this invention steadily climbs to around 0.75; subsequently, in the 11th round, the Dice score increases to above 0.90, with a peak value of approximately 0.92. In subsequent iterations, this core metric consistently remains in the high range above 0.90, demonstrating that this invention possesses a refined representation capability for irregular, diffuse smoke screen contours.

[0060] Figure 5 The variation trend of validation set mIoU with training cycles is shown. In the early training stage (Epochs 1-6), mIoU rises rapidly and exceeds 0.80, indicating that the model has initially achieved the extraction of shallow smoke features. In the middle training stage (Epochs 7-11), mIoU experiences a brief fluctuation before significantly improving, reaching around 0.95 by the 11th epoch, indicating that the network has further completed the fitting of deep and complex features. In the later training stage (Epochs 12-20), the curve gradually converges, and mIoU stabilizes in the high-precision range of 0.94 to 0.99. The overall evolution trend shows that the model has high feature extraction capability and good convergence stability in the dynamic smoke recognition task, and no obvious overfitting phenomenon is observed.

[0061] The evolution of the loss convergence state and validation set accuracy (Dice and mIoU) during the comprehensive training process shows that the method of this invention does not exhibit overfitting in the feature extraction and pixel-level classification tasks of dynamic smoke screen images. While achieving effective fitting of training samples, this invention demonstrates good generalization ability and perceptual robustness to unknown scene data, providing reliable underlying visual feature support for the analysis and accurate quantitative evaluation of the morphological evolution of dynamic smoke screens.

[0062] 2. Experimental results.

[0063] To intuitively verify the effectiveness of this method in dynamic smoke perception and recognition in complex real-world scenarios, cross-tests were conducted on dynamic smoke images from two different smoke-generating platforms: ground vehicles and drones.

[0064] Faced with large-scale, high-concentration, and highly irregularly shaped bursts of smoke generated by ground vehicles, this invention can accurately identify and delineate the diffuse outline of the smoke. Even under the interference of complex surface textures, it can clearly separate the smoke area from the background. In the scenario of dynamic smoke generation by UAVs in the air, for the more sparse, elongated, and diffused translucent smoke, as well as the highly interfering background of the sky, distant mountains, and complex terrain, the method demonstrates excellent pixel-level recognition accuracy, successfully achieving high-confidence separation of the smoke area from the background, without any missegmentation or omission.

[0065] The cross-scenario test results clearly demonstrate that this method can not only stably identify smoke screens of different shapes, but also possesses strong anti-interference capabilities and generalization robustness. Its high-quality, pixel-level smoke mask output accurately reproduces the true distribution pattern of smoke, perfectly adapting to different types of smoke generators and diverse application scenarios. This provides reliable underlying visual data support for the accurate quantitative evaluation of smoke generator operating status monitoring, smoke coverage area evolution, and concentration distribution patterns.

[0066] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for dynamic smoke screen image recognition for smoke generators, characterized in that: Includes the following steps: Step 1: Collect dynamic smoke screen images during the operation of the smoke generator, preprocess the collected images, and convert the collected color images into grayscale images; Step 2: Using the DPAUNet model, combined with deformable convolution and parameter-free attention mechanisms, extract deep latent features from the preprocessed dynamic smoke image and filter and optimize them, removing redundant features; including the following steps: Step 200: Construct the DPAUNet model; Step 201: Input the grayscale image preprocessed in Step 1 into the input layer of the DPAUNet model. The image is processed by the encoder bottleneck block in sequence to gradually extract the shallow detail features and deep semantic features of the dynamic smoke screen. Step 202: After the bottleneck block of each encoder layer is processed, the importance score of each pixel in the feature map is calculated by the SimAM parameterless attention module. Key smoke features and background interference features are selected to obtain the optimized feature map of each layer. The optimized feature maps of each layer of the encoder are passed to the corresponding layer of the decoder through skip connections. Step 3: Introduce a weighting factor and logarithmic transformation into the LogSAD algorithm to calculate the feature similarity between the image to be identified and the smoke template image. Compare this feature similarity with the recognition threshold to determine whether the image to be identified is a dynamic smoke screen image generated by the smoke generator.

2. The dynamic smoke screen image recognition method for smoke generators according to claim 1, characterized in that: Step 1 includes the following steps: Step 100: Acquire dynamic smoke screen images during the operation of the smoke generator using an image acquisition device. The acquired images are color RGB images, and the image sequence is stored in a standard image format. Step 101: Use Gaussian filtering to denoise the acquired color image; Step 102: Convert the denoised color image to a grayscale image; Step 103: Normalize the grayscale image and map it to the [0, 1] interval.

3. The dynamic smoke screen image recognition method for smoke generators according to claim 2, characterized in that: Step 101 includes the following steps: First, adjust the standard deviation σ of the Gaussian filter based on the noise intensity of the acquired image; Then, determine the size of the Gaussian filter convolution kernel; Finally, the convolution kernel is convolved with each pixel of the acquired color image, and the pixel value G(x,y) of the denoised dynamic smoke image at coordinates (x,y) is calculated according to the Gaussian filtering formula: ， Where (u,v) are the center coordinates of the Gaussian filter convolution kernel, and f(x,y) is the pixel value of the original smoke image at coordinates (x,y) before denoising; In step 102, a weighted average method is used for grayscale processing. The calculation formula for grayscale processing is as follows: Gray(x,y)=R1×R(x,y)+G1×G(x,y)+B1×B(x,y), Where Gray(x,y) is the gray value of the image at coordinates (x,y) after grayscale conversion, R(x,y), G(x,y), and B(x,y) are the pixel values ​​of the red, green, and blue channels of the denoised color image at coordinates (x,y), respectively, and R1, G1, and B1 are the weight coefficients of the red, green, and blue channels, respectively. Step 103 includes the following steps: Traverse the grayscale image and find the minimum pixel value G in the grayscale image. min and the maximum pixel value G max Then, according to the normalization formula, each pixel value Gray(x,y) is converted into a normalized pixel value G. norm (x, y), the calculation formula is as follows: .

4. The dynamic smoke screen image recognition method for smoke generators according to claim 1, characterized in that: The formula for calculating the bottleneck block features of the DPAUNet model is as follows: (1) The formula for 3×3 convolution feature compression is: F 3×3 =Conv 3×3 (X in ,W 3×3 ,b 3×3 ), In the formula, X in W is the input feature map for the bottleneck block. 3×3 b 3×3 The weights and biases of the 3×3 convolution are respectively, F 3×3 The compressed feature map output by the 3×3 convolution, Conv 3×3 This represents a standard convolution operation with a 3×3 dimension. (2) The adaptive capture formula for deformable convolution V2 features is: F dconv =DConv V2 (F 3×3 ,W d ,Δp,Δm), In the formula, W d F represents the convolution weights of deformable convolution V2, Δp is the adaptive offset, Δm is the offset weight, and F... dconv DConv outputs a deformable convolution V2 feature map. V2 This represents the deformable convolution V2 operation; (3) The formula for recovering 1×1 convolutional features is: F 1×1 =Conv 1×1 (F dconv ,W 1×1 ,b 1×1 ), In the formula, W 1×1 b 1×1 The weights and biases of the 1×1 convolution are respectively, F 1×1 Conv is the feature map of the recovered channel at the final output of the bottleneck block. 1×1 This represents a standard convolution operation of size 1×1.

5. The dynamic smoke screen image recognition method for smoke generators according to claim 1, characterized in that: In step 202, the importance scoring formula for the SimAM parameterless attention module is: ， In the formula, s i,j c represents the importance score of the j-th pixel in the i-th layer feature map; i,j Let μ be the pixel value of the j-th feature pixel in the i-th layer feature map. i Let v be the average pixel value of the feature map of the i-th layer. i Let σ be the global extremum of the pixel in the i-th layer feature map. i Let be the standard deviation of the pixels in the i-th layer feature map. To prevent the minimum value where the denominator is 0.

6. The dynamic smoke screen image recognition method for a smoke generator according to claim 5, characterized in that: In step 202, the formula for the feature map after filtering by the SimAM parameterless attention module is: F opt,i =Softmax(S i )·F enc,i , In the formula, F opt,i S is the optimized feature map of the i-th layer encoder after being filtered by the SimAM parameterless attention module. i The SimAM attention importance score for the feature map of layer i is determined by the sum of the values ​​of all pixels in that layer. i,j It is assembled according to position, and Softmax is a normalization operation. enc,i For the original feature map output by the i-th layer encoder, Softmax(S) i ) is for S i Perform normalization.

7. The dynamic smoke screen image recognition method for a smoke generator according to claim 6, characterized in that: In step 200, the DPAUNet model is a lightweight variant of UNet, whose core consists of an encoder, a decoder, deformable convolution V2, a SimAM parameterless attention module, and a skip connection structure. The DPAUNet model construction process is as follows: The input to the DPAUNet model is a preprocessed grayscale image; the encoder adopts a 4-layer lightweight bottleneck block structure, each bottleneck block consisting of a 3×3 convolutional layer, a deformable convolutional V2 layer, and a 1×1 convolutional layer; each bottleneck block is followed by a BatchNorm layer and a ReLU activation function. After the SimAM parameterless attention module is embedded in the bottleneck block of each layer of the encoder, no additional training parameters are required. By calculating the importance score of each pixel in the feature map, key smoke features are selected and background interference features are suppressed. The attention score is calculated based on the local contrast and global distribution of feature pixels. The decoder adopts a 4-layer deconvolution structure symmetrical to the encoder. At the same time, it fuses the shallow detail features extracted by the corresponding layer of the encoder with the deep semantic features of the decoder through skip connections. The output of the DPAUNet model is an optimized deep feature vector. The dimension of the feature vector is adaptively set according to the input image size and encoder structure. The output feature vector has completed the filtering of redundant features, retaining only the key features that can effectively represent the dynamic smoke screen.

8. The dynamic smoke screen image recognition method for a smoke generator according to claim 6, characterized in that: The training process of the DPAUNet model is as follows: Multiple sets of dynamic smoke screen images and non-smoke screen images from smoke generators are collected as training samples. The dynamic smoke screen images cover different types of smoke generators, different smoke concentrations, different diffusion stages, and different application scenarios. The non-smoke screen images cover background images similar to the application scenarios of the smoke generators, serving as negative samples. All training samples are preprocessed sequentially using step 1 to obtain standardized training sample images. Using the preprocessed training sample images as input and the sample labels as supervision signals, the DPAUNet model is trained using the cross-entropy loss function and the Adam optimizer. Iterative training continues until the model loss value converges and the recognition accuracy reaches a preset threshold, thus completing the training of the DPAUNet model.

9. The dynamic smoke screen image recognition method for a smoke generator according to claim 6, characterized in that: Step 3 includes the following steps: Step 300: Pre-collect multiple sets of dynamic smoke screen images during normal operation of the smoke generator as template samples; perform preprocessing in Step 1 and feature extraction and optimization using the DPAUNet model in Step 2 on each set of template samples to obtain the optimized feature vector of each set of template samples. Take the average of the optimized feature vectors of all template samples to obtain the baseline feature vector F of the smoke screen template. temp The baseline eigenvector F temp The dimension of the image to be recognized and the optimized feature vector F opt The dimensions are consistent; Step 301: Based on the importance score output by the SimAM parameterless attention module in step 202, calculate the optimized feature vector F of the image to be recognized. opt The weight factor w for each feature i ; Step 302, F opt F temp and w i Substitute the values ​​into the LogSAD algorithm formula to calculate the similarity value ILogSAD between the image to be identified and the template image; Step 303: Calculate the average similarity value between multiple sets of template feature vectors and use it as the recognition threshold T; Step 304: Compare the similarity value ILogSAD with the recognition threshold T to determine whether the image to be identified is a dynamic smoke screen image generated by a smoke generator. The judgment rules are as follows: If ILogSAD ≤ T, then the image to be identified is a dynamic smoke screen image generated by a smoke generator; If ILogSAD > T, then the image to be identified is determined to be either a non-smoke image or an abnormal smoke image.

10. The dynamic smoke screen image recognition method for a smoke generator according to claim 9, characterized in that: The formula for calculating the similarity value ILogSAD is as follows: ， In the formula, F opt,i F is the optimized feature vector of the image to be recognized. opt The i-th element, F temp,i It is the feature vector F after optimizing the smoke screen template image. temp The i-th element.

Images