A dust pollution source image enhancement processing method and system

Abstract

The present application relates to a kind of dust pollution source image enhancement processing method and system, specifically as follows: collecting dust pollution source image and environmental parameter, dust area and background are labeled to construct data set;With multiscale Gaussian kernel convolution estimation image illumination, reflection component and fusion are enhanced, then contour is strengthened by adaptive morphological operation, and morphological enhancement image is generated;Real dust texture is injected by random morphing field elastic transformation combined with texture synthesis, and synthesis enhancement image is obtained;Image enhancement model is constructed, input synthesis enhancement image, multiscale feature is extracted by encoder, decoder enhances feature and retains boundary, loss function is combined to optimize model, and the output result is also enhanced by region adaptive fusion, complete model training;Finally, new acquisition image is input into trained model, and the final enhanced dust pollution source image is generated.The present application improves dust image quality, accurately extracts feature, and helps pollution source identification and monitoring.

Description

Technical Field

[0001] This invention relates to the field of image enhancement technology, and in particular to a method and system for enhancing images of dust pollution sources. Background Technology

[0002] With the accelerating pace of urbanization, dust pollution has become increasingly serious, especially in environments such as construction sites, road construction sites, and dusty weather. Dust pollution sources not only affect air quality but also pose a threat to human health. To effectively monitor and control dust pollution, image processing-based pollution source detection technology has become an important research direction. Particularly in scenarios where images are acquired using drones or ground monitoring equipment, efficiently and accurately identifying and segmenting dust pollution sources has become a challenge in the field of image processing.

[0003] Chinese invention patent CN120746885A discloses a method for denoising and enhancing low-light physical examination images. The method comprises: acquiring a low-light physical examination image to be processed; obtaining a preliminary estimated illumination component of the low-light physical examination image based on a Retinex decomposition model; obtaining a precisely estimated illumination image based on the preliminary estimated illumination component; enhancing the precisely estimated illumination image using gamma transform to obtain an enhanced image; denoising the enhanced image using the BM3D algorithm to obtain a denoised image; and merging the enhanced image and the denoised image to obtain a denoised and enhanced image.

[0004] Existing technologies have several shortcomings in practical applications: conventional image enhancement methods often fail to effectively handle low contrast and blurred boundaries in dust pollution source images, and are prone to introducing noise during the enhancement process, failing to adaptively adjust the enhancement intensity for different regions; traditional image segmentation methods typically rely on single-scale convolution kernels, making it difficult to simultaneously capture both fine particles and macroscopic pollution areas of dust pollution sources, resulting in an inability to accurately segment the complex shapes of dust pollution sources; existing data augmentation methods, such as rotation and scaling, cannot simulate the varied shapes of real dust pollution sources, leading to insufficient diversity in training data, which may result in overfitting of the model and an inability to generalize to real-world scenarios; conventional loss functions are often insensitive to the segmentation results of blurred boundaries and irregular shapes, failing to effectively constrain the model's accurate boundary recognition, resulting in low segmentation accuracy.

[0005] Therefore, this invention proposes an image enhancement processing method and system for dust pollution sources to solve the above problems. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention develops a method and system for enhancing images of dust pollution sources. This invention can accurately extract features and precisely identify segmentation boundaries, thereby improving the quality of dust images.

[0007] The technical solution of this invention to solve the technical problem is a method for image enhancement processing of dust pollution sources, comprising the following steps:

[0008] S1. Collect images of dust pollution sources using on-site camera equipment, record corresponding environmental parameters, and annotate the images. The annotation categories are divided into dust pollution source areas and background areas to form a dust pollution source image dataset.

[0009] S2. The illumination and reflection components of the dust pollution source images in the dataset are estimated by multi-scale Gaussian kernel convolution. Then, the components at each scale are fused to obtain a multi-scale enhanced image. Adaptive morphological operations are used to enhance the contour information of the irregular shape of the multi-scale enhanced image to generate a morphological enhanced image.

[0010] S3. By generating a random deformation field, the morphologically enhanced image is elastically transformed, and real dust texture features are injected by combining texture synthesis technology to obtain a synthetic enhanced image.

[0011] S4. Construct a dust pollution source image enhancement model. Input the synthesized enhanced image into the model. First, extract multi-scale features through a staged downsampling encoder and output a fused multi-scale feature map. Then, enhance the pollution source features and retain boundary details through a decoder and output a skip connection fused feature. Calculate the loss function based on the corresponding predicted segmentation mask and the labeled real mask output by the encoder to optimize the model. Then, perform region adaptive fusion enhancement based on the fused multi-scale feature map, predicted segmentation mask, and attention mechanism to output an enhanced dust pollution source image.

[0012] The model is trained and optimized to obtain a well-trained image enhancement model for dust pollution sources;

[0013] S5. Input the newly acquired dust pollution source images into the trained model to generate the final enhanced dust pollution source images.

[0014] S1 is as follows:

[0015] The camera equipment is high-resolution, including fixed surveillance cameras and mobile drone platforms;

[0016] The collected data covers various forms under different time periods, weather conditions, and lighting environments, including diffuse, aggregated, and scattered forms, while also recording corresponding environmental parameters;

[0017] The images were finely annotated using a combination of manual annotation and semi-automatic tools. The dust pollution source area refers to the core area in the image that directly shows the diffusion of dust, smoke or particulate matter, while the background area refers to the non-polluting parts, including the sky, buildings and vegetation. During the annotation process, environmental experts guided the annotators to use image annotation software to delineate the pollution source boundaries pixel by pixel and generate a binary segmentation mask. In the mask, the pollution source area was marked as the foreground and the background area was marked as the background.

[0018] The labeled image data is used to construct a dataset, which is then divided into a training set, a validation set, and a test set.

[0019] S2 is as follows:

[0020] The illumination component of the original dust pollution source image is estimated by applying multi-scale Gaussian kernel convolution, and the reflection component is calculated in the logarithmic domain. Then, based on the fusion weights of different scales, the fusion enhancement result is adaptively fused to output a multi-scale enhanced image.

[0021] Then, by applying an adaptive structuring element based on local gradient magnitude to the multi-scale enhanced image, morphological opening is performed. The value of the structuring element is dynamically adjusted according to the gradient threshold to eliminate noise in the image and strengthen the boundaries of irregular shapes, thus generating a morphologically enhanced image.

[0022] S3 is as follows:

[0023] A smooth random deformation field is generated based on radial basis function interpolation, and the deformation offset is calculated by random offset weights of control points. By applying non-rigid transformation to the morphologically enhanced image, the irregular shape changes of dust pollution sources are simulated, generating a deformation-enhanced image, thereby increasing the shape diversity of training samples.

[0024] By randomly selecting texture blocks from a dust texture library and extracting high-frequency details of the textures based on the Laplacian operator, and combining the texture fusion intensity, the texture blocks are seamlessly integrated into the deformation enhancement image to obtain a synthetic enhancement image.

[0025] The specific operations in the encoder are as follows:

[0026] The encoder body is a staged serial downsampling structure, containing L stages, each stage consisting of convolution, activation functions and pooling layers;

[0027] Each stage extracts features step by step through a series of convolution, activation functions and pooling operations. The input of the first stage is the synthesized enhanced image, and the output of the first stage is used as the input of the second stage. Similarly, the output of the previous stage is used as the input of the next stage.

[0028] Then, the output of each stage is passed through multiple parallel dilated convolution branches, with the number of branches corresponding to the number of stages. Each branch uses convolution kernels with different dilation rates to perform dilated convolution operations on the output of the last stage of the encoder, and combines them with the modified linear unit activation function to generate the output feature map.

[0029] Finally, the output feature maps of each branch dilated convolution are upsampled to adjust the spatial size, and weighted fusion is performed by combining adaptive fusion weights and attention masks to generate fused multi-scale feature maps.

[0030] The specific operations in the decoder are as follows:

[0031] A spatial attention mechanism is used to generate an attention weight map. The fused multi-scale feature map is deconvolved and upsampled. An edge enhancement mask is combined to highlight the boundary region. The pollution source-related region in the feature map is dynamically weighted to enhance the feature response of the pollution source region and generate an attention-weighted upsampled feature map. The edge enhancement mask is generated based on the gradient magnitude of the multi-scale enhanced image.

[0032] Then, the output feature maps of different stages of the encoder are convolved and adjusted, and then skipped connections are made with the attention-weighted upsampled feature map. The feature maps are then fused with adaptive weights and edge enhancement attention maps to retain fine-grained boundary information and generate skip connection fused feature maps. The edge enhancement attention map is generated based on the gradient magnitude of the multi-scale enhanced image.

[0033] Design a composite loss function that combines shape-aware terms, boundary distance constraints, and cross-entropy loss. The calculation process of the loss function is as follows:

[0034] First, the skip connection fusion feature map output by the decoder is processed... The predicted segmentation mask is obtained by processing convolution and sigmoid activation function;

[0035] The compactness ratio is calculated based on the region area and boundary perimeter of the predicted segmentation mask and the labeled real segmentation mask, and the empty mask is processed by the indicator function, and then the shape-aware loss term is calculated.

[0036] By calculating the Hausdorff distance between the predicted boundary and the true boundary, and taking the maximum value of the two-way distance as the boundary alignment loss, the consistency between the predicted boundary and the true boundary is constrained, and then the boundary alignment loss term is calculated.

[0037] The difference between the predicted probability and the true label of each pixel is calculated based on the predicted segmentation mask and the labeled true segmentation mask, and the cross-entropy loss, which measures the pixel-level classification error, is obtained.

[0038] The total loss function is calculated by weighting and summing the cross-entropy loss, shape-aware loss, and boundary alignment loss.

[0039] The specific process of region adaptive fusion enhancement is as follows:

[0040] Multi-scale detail enhancement of the pollution source region is performed by using the encoder's multi-scale feature map and the predicted segmentation mask. By calculating the multi-scale detail map and weighted fusion, the detail visibility of the pollution source region is enhanced, and a multi-scale detail-enhanced image is obtained.

[0041] By using illumination component estimation and texture synthesis strategies, the background area is processed to homogenize the illumination and inject realistic dust texture to generate a background fusion image;

[0042] Based on spatial attention weight maps and edge enhancement masks, adaptive weight fusion is performed on multi-scale detail enhancement images and background fusion images to generate the final enhanced dust pollution source image.

[0043] The model training process is as follows:

[0044] The dataset is divided into training, validation, and test sets. The model is iteratively trained using the data in the training and validation sets. First, the trainable parameters of the encoder and decoder modules are initialized, the weights are initialized using a random normal distribution, and the optimizer, initial learning rate, and batch size are set.

[0045] In each iteration, the batch of synthesized enhanced images and their corresponding real segmentation masks are input, the predicted segmentation mask is calculated through forward propagation, and the difference between the model output and the real segmentation mask is evaluated based on the loss function.

[0046] During backpropagation, gradients are calculated and model parameters are updated using the chain rule to minimize the loss function, while gradient pruning is employed to prevent gradient explosion.

[0047] Iterative training monitors performance on the validation set. After each round of training, the loss and segmentation accuracy metrics on the validation set are calculated to evaluate the model's generalization ability.

[0048] The criteria for stopping iteration are set based on the early stopping mechanism and the maximum number of iterations;

[0049] After training, save the parameters of the model that performs best on the validation set to obtain the trained model;

[0050] The trained model was tested using data from the test set.

[0051] This invention also proposes an image enhancement processing system for dust pollution sources, which performs an image enhancement processing method for dust pollution sources, including:

[0052] Data acquisition module: Acquires images of dust pollution sources through camera equipment, records corresponding environmental parameters, annotates the acquired images, and forms a dust pollution source image dataset;

[0053] Data preprocessing module: includes contour enhancement unit and composite enhancement unit;

[0054] The system receives data from the dust pollution source image dataset in the data acquisition module, passes it sequentially through the contour enhancement unit and the synthesis enhancement unit, preprocesses the data in the dataset, and outputs the preprocessed data.

[0055] Dust pollution source image enhancement module: includes a dust pollution source image enhancement unit and an optimization unit;

[0056] The system receives preprocessed data from the data preprocessing module, then outputs enhanced dust pollution source images through the dust pollution source image enhancement unit. Finally, the optimization unit optimizes the output of the dust pollution source image enhancement unit.

[0057] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects:

[0058] This invention addresses the problem of traditional image enhancement methods' inability to adaptively adjust contrast in different regions by combining multi-scale illumination estimation with reflectance component extraction. This effectively avoids excessive noise enhancement and loss of detail. Furthermore, by fusing Gaussian kernels of different scales, it achieves a balance between detail enhancement and noise suppression, improving the visibility and detail rendering of dust pollution source images. This invention also introduces adaptive morphological operations, adjusting morphological structural elements through local gradient magnitudes to enhance the contour information of irregular pollution sources. This effectively highlights the boundaries of dust pollution sources and avoids false noise enhancement when processing blurry or low-contrast images, ensuring the clarity and accuracy of boundaries. This invention employs a random deformation field generation and elastic transformation strategy based on radial basis function interpolation to simulate the irregular shape changes of dust pollution sources, thereby increasing the shape diversity of the training dataset. In addition, it combines texture synthesis technology to inject realistic dust textures, making the textures of the training samples more realistic and improving the model's adaptability to pollution sources in real-world scenarios. This invention uses a composite loss function based on shape-aware terms and boundary distance constraints to solve the problem of poor adaptability of traditional loss functions to fuzzy boundaries and irregular shapes of dust pollution sources. By comprehensively considering shape compactness and boundary alignment, it can effectively improve the model's segmentation accuracy and reduce shape errors and boundary inconsistencies. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0060] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0061] Figure 2 Images of dust pollution sources were collected.

[0062] Figure 3 Pixel value normalization feature map for morphological enhancement of images.

[0063] Figure 4 The encoder consists of feature maps for three stages, where... Figure 4-1 This represents the high-resolution feature map output from encoder stage 1. Figure 4-2 This represents the medium-resolution feature map output from encoder stage 2. Figure 4-3 This represents the low-resolution feature map output from encoder stage 3.

[0064] Figure 5 The output feature maps are the three dilated convolution branches, where... Figure 5-1 This represents the feature map output by dilated convolution branch 1. Figure 5-2 This represents the feature map output by dilated convolution branch 2. Figure 5-3 This represents the feature map output by dilated convolution branch 3.

[0065] Figure 6 A graphical representation of the thermal values ​​corresponding to the boundary of a dust pollution source.

[0066] Figure 7 This is a schematic diagram for predicting the segmentation mask.

[0067] Figure 8 A normalized heatmap of the region with boundary distance.

[0068] Figure 9 Pixel value normalization feature maps are used to enhance the image's detail at multiple scales.

[0069] Figure 10 To enhance the image of dust pollution sources. Detailed Implementation

[0070] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0071] Example 1

[0072] like Figure 1 As shown, a method for enhancing images of dust pollution sources includes the following steps:

[0073] S1. Collect images of dust pollution sources using on-site camera equipment, record corresponding environmental parameters, and annotate the images. The annotation categories are divided into dust pollution source areas and background areas to form a dust pollution source image dataset.

[0074] S2. The illumination and reflection components of the dust pollution source images in the dataset are estimated by multi-scale Gaussian kernel convolution. Then, the components at each scale are fused to obtain a multi-scale enhanced image. Adaptive morphological operations are used to enhance the contour information of the irregular shape of the multi-scale enhanced image to generate a morphological enhanced image.

[0075] S3. By generating a random deformation field, the morphologically enhanced image is elastically transformed, and real dust texture features are injected by combining texture synthesis technology to obtain a synthetic enhanced image.

[0076] S4. Construct a dust pollution source image enhancement model. Input the synthesized enhanced image into the model. First, extract multi-scale features through a staged downsampling encoder and output a fused multi-scale feature map. Then, enhance the pollution source features and retain boundary details through a decoder and output a skip connection fused feature. Calculate the loss function based on the corresponding predicted segmentation mask and the labeled real mask output by the encoder to optimize the model. Then, perform region adaptive fusion enhancement based on the fused multi-scale feature map, predicted segmentation mask, and attention mechanism to output an enhanced dust pollution source image.

[0077] The model is trained and optimized to obtain a well-trained image enhancement model for dust pollution sources;

[0078] S5. Input the newly acquired dust pollution source images into the trained model to generate the final enhanced dust pollution source images.

[0079] In a specific implementation, S1 is as follows:

[0080] High-resolution camera equipment deployed on-site systematically collects images of dust pollution sources. The equipment includes fixed monitoring cameras and mobile drone platforms, covering different time periods, weather conditions, and lighting environments to ensure data diversity.

[0081] The data collection process focuses on capturing various forms of dust pollution sources in real-world scenarios, such as diffuse, aggregated, and drifting forms. It also records environmental parameters corresponding to the images, such as particulate matter concentration and wind speed, for subsequent analysis.

[0082] The collected images of dust pollution sources were initially screened to remove severely distorted or irrelevant samples, retaining images with typical dust characteristics as basic data.

[0083] In the construction of the training dataset, a combination of manual annotation and semi-automatic tools was used to perform fine annotation on the images. The annotation categories were strictly divided into two categories: dust pollution source areas and background areas. Dust pollution source areas refer to the core areas in the image that directly show the diffusion of dust, smoke or particulate matter, while background areas include non-polluted parts such as the sky, buildings, and vegetation.

[0084] During the annotation process, environmental experts guide the annotators to use image annotation software to delineate the boundaries of pollution sources pixel by pixel and generate a binary segmentation mask. In the mask, the pollution source area is marked as the foreground and the background area is marked as the background to ensure the accuracy and consistency of the annotation.

[0085] The constructed dataset contains thousands of high-quality images and their corresponding masks, and is divided into training, validation and test sets according to proportions;

[0086] During the specific implementation process, images of dust pollution sources were collected, such as... Figure 2 As shown.

[0087] In a specific implementation, S2 is as follows:

[0088] Images of dust pollution sources suffer from severe blurring and low contrast due to atmospheric scattering, particulate matter diffusion, and uneven illumination, resulting in unclear boundaries between the pollution source area and the background. Conventional processing methods such as histogram equalization or single-scale enhancement often over-amplify noise or fail to adaptively adjust the contrast of different regions, making it difficult to effectively recover key details of blurred boundaries. Therefore, this invention enhances the contour information of irregular shapes by estimating illumination and reflection components at multiple scales and combining them with adaptive morphological operations. The specific steps are as follows:

[0089] 1) Multiscale illumination estimation and reflection component extraction

[0090] Illumination components are estimated from the original dust pollution source image by applying multi-scale Gaussian kernel convolution, and reflectance components are calculated in the logarithmic domain. Then, the fusion enhancement result is adaptively enhanced based on fusion weights of different scales to balance detail enhancement and noise suppression, thereby improving the visibility and detail of the pollution source area. This is expressed as follows:

[0091]

[0092] In the formula, Indicates enhanced image exist The intensity value at the pixel location is used to improve the visibility and detail of the pollution source area;

[0093] Image representing the original dust pollution source exist The intensity value at the pixel location, due to atmospheric scattering, particulate matter diffusion and uneven lighting, results in an overall blurred image, low contrast, unclear boundary between the pollution source area and the background, and may contain noise.

[0094] Indicates the first The fusion weights at each scale satisfy... The preferred setting value is To emphasize details at medium scales and balance noise suppression;

[0095] Indicates the first A Gaussian kernel function with scales and a standard deviation of . , used to simulate light components;

[0096] Indicates pixel coordinate index;

[0097] Indicates the first The standard deviation of each Gaussian kernel, with a value sequence of... Corresponding to large, medium, and small scales, it controls the smoothness of convolution;

[0098] For scale indexing, it is preferable to set the value range to 1 to 3;

[0099] This represents the total number of scales, with a preferred value of 3, representing the number of branches in multi-scale fusion.

[0100] This represents a two-dimensional convolution operation;

[0101] This represents a logarithmic function, with the default base being the natural constant.

[0102] It should be noted that the fusion weight The settings need to balance detail enhancement and noise suppression. A weight of 0.5 corresponds to a medium scale, emphasizing key details; a weight of 0.3 corresponds to a large scale, used to suppress noise; and a weight of 0.2 corresponds to a small scale, preserving subtle features. The standard deviation of the Gaussian kernel is set to the sequence... It corresponds to three scales: small, medium, and large. The preset values ​​are based on the multi-scale characteristics of dust pollution source images. Small scale captures fine particles, medium scale enhances the outline of pollution sources, and large scale processes global illumination unevenness.

[0103] 2) Adaptive morphological contour enhancement

[0104] By applying morphological opening operations to the enhanced image using an adaptive structuring element based on local gradient magnitude, and dynamically adjusting the value of the structuring element according to the gradient threshold, noise is eliminated and the boundaries of irregular shapes are enhanced, highlighting the irregular contours of pollution sources. This is represented as:

[0105]

[0106] In the formula, Representing morphologically enhanced images exist The intensity value at a pixel location is used to highlight the irregular contours of the pollution source, morphologically enhancing the image as follows. Figure 3 As shown;

[0107] This represents the morphological opening operation, used to eliminate noise and preserve boundary structure;

[0108] Indicates the adaptive binary structuring element in The value at the pixel location is dynamically generated based on the local gradient and is used to adjust the sensitivity of morphological operations.

[0109] In its implementation, the adaptive binary structuring element is generated based on the local gradient magnitude. Specifically, if... ,but ,like ,but 0;

[0110] in, This indicates an enhanced image of dust pollution sources. exist Gradient components along the axial direction;

[0111] This indicates an enhanced image of dust pollution sources. exist Gradient components along the axial direction;

[0112] This represents the gradient threshold, and the preferred value is [value missing]. This is used to distinguish between high-gradient boundary regions and low-gradient flat regions, in order to adaptively adjust the structuring elements. The maximum gradient magnitude in the image is represented by 10%. Taking 10% as the threshold can effectively distinguish between high gradient boundary regions such as pollution source outlines and low gradient flat regions, thus avoiding noise amplification.

[0113] Represents the maximum value function. The term indicates the return matrix. The maximum element value in;

[0114] This indicates an enhanced image of dust pollution sources. The gradient vector field, i.e. ;

[0115] This represents the L2 norm, also known as the Euclidean norm.

[0116] This indicates the transpose operation.

[0117] In a specific implementation, S3 is as follows:

[0118] Because dust pollution sources are highly irregular in shape and randomly distributed in space, conventional data augmentation methods such as rotation and scaling cannot fully simulate the morphological diversity of pollution sources in real-world scenes, leading to model overfitting to a limited number of training samples. Therefore, this invention generates a random deformation field to elastically transform the preprocessed image and combines this with texture synthesis technology to inject realistic dust texture features, thereby increasing the shape diversity and texture realism of the training samples. The specific steps are as follows:

[0119] 1) Generation of random deformation fields and elastic transformation

[0120] A smooth random deformation field is generated based on radial basis function interpolation, and the deformation offset is calculated using the random offset weights of control points. A non-rigid transformation is applied to the morphologically enhanced image to simulate the irregular shape changes of dust pollution sources, increasing the shape diversity of the training samples. This is represented as follows:

[0121]

[0122] In the formula, Images representing sources of dust pollution with enhanced deformation exist The intensity value at the pixel location is used to increase the shape diversity of the training samples;

[0123] Represents morphologically enhanced dust images exist Intensity value at pixel location;

[0124] Indicates the location Place The deformation offset in the axial direction is calculated as follows: ;

[0125] Indicates the location Place The deformation offset in the axial direction is calculated as follows: ;

[0126] Indicates the number of control points, with values ​​ranging from 1 to 2. , used to generate deformation fields;

[0127] Indicates the first The coordinates of each control point are evenly distributed on the image grid, serving as deformation reference points;

[0128] This represents the index of the control point, with a value ranging from 1 to... ;

[0129] Indicates the first Each control point is The random offset weights along the axis are preferably set to have a mean of 0 and a standard deviation of 0. The weights of random sampling in the normal distribution of pixels are used to generate deformation offsets to simulate random deformation;

[0130] Indicates the first Each control point is The random offset weights along the axis are preferably set to have a mean of 0 and a standard deviation of 0. Weights randomly sampled from the normal distribution of pixels are used to generate deformation offsets to simulate random deformation;

[0131] This represents the radial basis function, used to smooth the interpolated deformation field. Preferably, the scale parameter of the radial basis function is set to... Pixel.

[0132] It should be noted that control points are reference points evenly distributed on the image grid, used for deformation interpolation. The more control points there are, the more refined the deformation, which can simulate the irregular shape changes of dust pollution sources.

[0133] 2) Dust texture synthesis and fusion

[0134] By randomly selecting texture blocks from a dust texture library and extracting high-frequency details of the texture based on the Laplacian operator, and then seamlessly integrating the texture blocks into the deformed image by combining texture fusion intensity, the texture realism of the pollution source area is enhanced. This results in a synthesized image with higher shape diversity and texture realism, as shown below:

[0135]

[0136] In the formula, Indicates synthetic enhanced image exist The intensity value at the pixel location, combined with the deformed image and the realistic dust texture, has higher shape diversity and texture realism;

[0137] This represents the texture blending intensity, with a preferred value of [value to be filled in]. This controls the extent to which textures are added;

[0138] This represents the Laplacian operator, used to extract high-frequency details from textures;

[0139] This indicates that a texture block was randomly selected from the dust texture library. The intensity value at a pixel location is a local texture block randomly cropped from a library of real dust pollution source images, containing typical dust textures.

[0140] It should be noted that, The term is used as a Laplacian operator for texture blocks to extract high-frequency details, enhance the naturalness of texture blending, and enable the added texture to be seamlessly integrated with the deformed image.

[0141] S4 is as follows:

[0142] In a specific implementation, the operations in the encoder are as follows:

[0143] Dust pollution sources exhibit multi-scale characteristics in images, and due to ambiguity, detailed features are expressed inconsistently at different resolutions. Conventional fully convolutional network encoders use single-scale convolutional kernels, making it difficult to capture both fine particles and macroscopic pollution areas simultaneously.

[0144] This invention expands the receptive field without sacrificing resolution by deploying multi-branch dilated convolutional layers and a feature pyramid network in parallel within the encoder section. It also combines top-down feature fusion to aggregate multi-scale information and adds a multi-scale feature fusion module at the encoder end to further capture multi-scale contextual information. This allows for the simultaneous capture of both fine particles and macroscopic contamination areas. The specific steps are as follows:

[0145] 1) Encoder staged encoding

[0146] The encoder body is a staged serial downsampling structure, containing L stages, each stage consisting of convolution, activation functions and pooling layers;

[0147] in, This represents the number of encoder stages, i.e., the number of downsampling layers in the encoder, with a preferred value of [value to be filled in]. Each stage outputs feature maps at different resolutions.

[0148] definition Indicates the encoder's first... The feature map of the stage is used to inject detailed information. This is the encoder stage index, with a value range of [value range missing]. arrive ;

[0149] In one implementation, the encoder structure consists of three stages. Each stage downsamples and extracts features from the input image through convolutional and pooling layers, outputting feature maps at different resolutions. Specifically, the encoder extracts features step-by-step through a series of convolutional, activation function, and pooling operations, such as... Corresponding to stage 1, the input synthetically enhanced dust pollution source image Convolution and pooling are performed to output a high-resolution feature map, which is the feature map of the encoder's first stage. , Corresponding to stage 2, for Further downsampling is performed to output a medium-resolution feature map, which is the feature map of the second stage of the encoder. , Corresponding to stage 3, for Downsampling is performed to output a low-resolution feature map, which is the feature map of the encoder's third stage. These feature maps can capture multi-scale information from details to the global scale of dust pollution source images and are fused with decoder features in skip connections to inject fine-grained boundary information and improve segmentation accuracy.

[0150] 2) Multi-scale feature fusion module

[0151] By applying three parallel dilated convolution branches, each using convolution kernels with different dilation rates to perform dilated convolution operations on the feature maps of the third stage of the encoder, and combining this with a modified linear unit activation function to generate output feature maps, the receptive field is expanded without reducing resolution, capturing dust features at different scales. This is represented as follows:

[0152]

[0153] In the formula, Indicates the first The output feature maps of each dilated convolution branch are used to capture dust features at different scales.

[0154] The branch index of the dilated convolution has a value range of 1. arrive ;

[0155] Indicates the first The convolution kernel weight matrix of each branch has a size of [size missing]. Used for feature extraction;

[0156] The number of input channels should correspond to the feature map of the third stage of the encoder. The number of channels is the same;

[0157] This represents the number of output channels, with a preferred value of [value to be filled in]. Control the number of feature map channels;

[0158] Indicates the void ratio The dilated convolution operation is used to expand the receptive field without reducing the resolution;

[0159] Indicates the first The dilation rate of each dilated convolution branch, with a value sequence of... This controls the sampling interval of the convolution kernel;

[0160] Indicates the first The bias term of each dilated convolution branch is used to adjust the output features;

[0161] This represents the modified linear unit activation function.

[0162] 3) Feature Pyramid Fusion

[0163] By upsampling the output feature map of multi-branch dilated convolution to adjust its spatial size, and then combining it with adaptive fusion weights and attention masks for weighted fusion, a fused feature map with multi-scale receptive fields is generated, thereby simultaneously capturing features of fine particles and macroscopic contamination regions, as shown below:

[0164]

[0165] In the formula, This represents the fusion of multi-scale feature maps, with dimensions of [missing information]. It has a multi-scale receptive field, which can simultaneously capture the characteristics of fine particles and macro-polluted areas, improving the accuracy and robustness of dust pollution source segmentation.

[0166] Indicates the first The adaptive fusion weights at each scale are calculated using global average pooling and a fully connected layer, and are represented as follows: This is used to balance contributions at different scales;

[0167] Indicates the first The logarithmic score of the adaptive fusion weights at each scale is calculated as follows: ;

[0168] Indicates the first Log score of adaptive fusion weights at each scale;

[0169] Indicates difference from The branch index of the dilated convolution has a value range of 1. arrive ;

[0170] Indicates the first The fully connected layer weight vectors of scale n, with dimension n. This is used to map global average pooling features to scalar scores;

[0171] Indicates the first A fully connected layer bias term at each scale is used to adjust the score;

[0172] This indicates a global average pooling operation. Xiang Jiangdi Output feature map of each dilated convolution branch Spatial dimension pooling Preserve channel dimension, output dimension is ;

[0173] Indicates an upsampling operation. The term represents the first term. Output feature map of each dilated convolution branch Perform an upsampling operation to make its spatial size consistent with other branches, i.e., the spatial size is... The upsampling method uses bilinear interpolation.

[0174] This represents element-wise multiplication;

[0175] Indicates the first Attention masks for each dilated convolutional branch are used for feature weighting to highlight important regions, and are calculated as follows: ;

[0176] Indicates the first The convolutional weights of the mask generation layer are used to learn the attention map and are trainable parameters.

[0177] Indicates the first The bias of each mask generation layer, used to adjust the output, is a trainable parameter.

[0178] This represents the Sigmoid activation function;

[0179] The height of the feature map is represented by the original dust pollution source image. The heights are the same;

[0180] The width of the feature map is compared to the original dust pollution source image. They have the same width.

[0181] The feature maps output by the encoder in the three stages are as follows: Figure 4 As shown, the feature maps output by the three parallel dilated convolution branches are as follows: Figure 5 As shown.

[0182] In a specific implementation, the decoder operates as follows:

[0183] The irregular shape of dust pollution sources makes it easy to lose detailed information during the upsampling process of the decoder. Conventional fully convolutional network decoders use simple deconvolution and cannot adaptively recover key boundaries.

[0184] This invention enhances the feature response of pollution source regions while preserving fine-grained boundary information by dynamically weighting the feature map of pollution source-related regions through a spatial attention module integrated into the decoder, and by injecting multi-scale details into the encoder using improved skip connections. The specific steps are as follows:

[0185] 1) Spatial attention weighted upsampling

[0186] An attention weight map is generated using a spatial attention mechanism. This map is then deconvolved and upsampled onto the fused multi-scale feature map. An edge enhancement mask is then used to highlight boundary regions, thereby dynamically weighting the pollution source-related regions in the feature map and enhancing the feature response of these regions. This is represented as follows:

[0187]

[0188] In the formula, This represents an attention-weighted upsampling feature map, used to enhance the feature response of the pollution source region;

[0189] Representation of spatial attention weights exist The intensity value at a pixel location is used to guide feature selection, and its calculation method is expressed as follows: ;

[0190] This indicates the deconvolution operation. The term represents the fusion of multi-scale feature maps. Deconvolutional upsampling is performed to restore spatial resolution and generate high-resolution features. Item indicates to The edge-enhanced fusion feature map generated by the item is deconvolutionally upsampled to highlight the boundary region;

[0191] Represents the fusion of multi-scale feature maps exist Intensity value at pixel location;

[0192] This represents an edge enhancement mask, generated based on gradient magnitude, and calculated as follows: High-value areas correspond to the boundaries of dust pollution sources;

[0193] This represents the convolutional weight matrix of the edge mask generation layer, with size . , used to learn edge attention from gradient magnitude, is a trainable parameter;

[0194] This represents the bias term of the edge mask generation layer, used to adjust the output, and is a trainable parameter.

[0195] The weight matrix of the attention convolutional layer is used to generate the attention map and is a trainable parameter.

[0196] This represents the bias term of the attention convolutional layer, used to adjust the output, and is a trainable parameter.

[0197] The visualization of the thermal values ​​corresponding to the boundaries of dust pollution sources obtained based on edge enhancement masks is shown below. Figure 6 As shown.

[0198] 2) Multi-scale skip connection fusion

[0199] By convolving and adjusting the feature maps from different stages of the encoder and then performing skip connections with the decoder features, and combining them with adaptive weights and edge enhancement attention maps, fine-grained boundary information is preserved, thereby improving the integrity of the dust pollution source outline. This can be represented as follows:

[0200]

[0201] In the formula, This represents a skip connection fusion feature map, used to combine decoder upsampling features and encoder multi-scale features to inject fine-grained boundary information and improve the integrity of dust pollution source contours;

[0202] Indicates the first The adaptive weights of each skip connection are calculated using global average pooling and a fully connected layer, and are represented as follows: This is used to balance the contributions of different encoder stages;

[0203] Indicates the first The fully connected layer weight vector, which consists of adaptive weights for each skip connection, is used to map global average pooling features to scalar scores and is a trainable parameter.

[0204] Indicates the first The fully connected layer bias term, which is an adaptive weight of each skip connection, is used to adjust the score and is a trainable parameter.

[0205] Indicates the first Edge enhancement attention maps in each encoder stage help the model better capture boundary information and reduce detail loss in blurry and irregularly shaped dust pollution source images. The calculation method is as follows: ;

[0206] Indicates the first The convolutional weight matrix of the edge mask generation layer in each encoder stage is used to learn edge attention from gradient magnitude and is a trainable parameter.

[0207] Indicates the first The bias term of the edge mask generation layer in each encoder stage is used to adjust the output and is a trainable parameter.

[0208] Indicates the first A jump connection The convolution kernel, used for channel adjustment, is a trainable parameter;

[0209] This indicates a skip connection operation. If the feature sizes of the skip connections are the same, then an element-wise addition operation is performed directly. If the feature sizes of the skip connections are different, then a bilinear interpolation upsampling operation is used to upsample the encoder's first element. Stage feature map After upsampling, its spatial size is compared with the attention-weighted upsampled feature map. If they match, then perform element-by-element addition.

[0210] In a specific implementation, the calculation process of the loss function is as follows:

[0211] In dust pollution source segmentation, fuzzy boundaries and irregular shapes prevent the cross-entropy loss function from accurately penalizing boundary errors, while conventional loss functions such as Dice loss are not sensitive to shape changes.

[0212] This invention optimizes the model by designing a composite loss function that combines shape-aware terms and boundary distance constraints, calculating shape compactness and boundary alignment. This encourages predicted segmentation masks to approximate compact shapes and constrains the consistency between predicted and true boundaries. The specific steps are as follows:

[0213] 1) Shape-aware loss calculation

[0214] The compactness ratio is calculated based on the region area and boundary perimeter of the predicted segmentation mask and the real mask. An indicator function is used to process empty masks, thereby measuring shape regularity and encouraging the predicted segmentation mask to approximate a compact shape, as expressed below:

[0215]

[0216] In the formula, This represents the shape-aware loss term, used to encourage the predicted segmentation mask to approximate a compact shape;

[0217] This represents the predicted segmentation mask, which is the final predicted segmentation mask output by the model. It is formed by the skip connection fused feature map output by the decoder. go through The binarized or probabilistic map obtained by convolution and Sigmoid activation function processing represents the predicted area of ​​dust pollution source;

[0218] To prevent division by zero decimals and to prevent the denominator from being 0, it is preferable to set it to 0.00001.

[0219] The mask represents the true segmentation, which is a binary image. The foreground region represents the dust pollution source, and the background is the unpolluted region.

[0220] This represents the function for calculating the area of ​​the mask region, used to quantize the mask size;

[0221] This represents the function for calculating the perimeter of the mask boundary, used to quantize the boundary length;

[0222] Pi is a constant.

[0223] This is an indicator function that takes the value when the condition is true. Otherwise , used to handle null masks.

[0224] It should be noted that in the binary mask, the foreground is represented as white, corresponding to the value 1 in the binary representation. A schematic diagram of the prediction segmentation mask is shown below. Figure 7 As shown.

[0225] 2) Boundary distance constraints

[0226] By calculating the Hausdorff distance between the predicted and true boundaries, the maximum value of the two-way distance is taken as the boundary alignment loss to constrain the consistency between the predicted and true boundaries and optimize the model's penalty for boundary errors. The region-normalized heatmap of the boundary distance is shown in Figure 8, and the calculation formula is as follows:

[0227]

[0228] In the formula, This represents the boundary alignment loss term, used to constrain the consistency between the predicted boundary and the true boundary;

[0229] This indicates taking the maximum of two values, used to calculate the bidirectional Hausdorff distance;

[0230] Represents the predicted segmentation mask The set of boundary points;

[0231] Represents the true segmentation mask The set of boundary points;

[0232] Represents the predicted segmentation mask The coordinate index of the boundary point set;

[0233] Represents the true segmentation mask The coordinate index of the boundary point set;

[0234] This represents the supremum operation, used to find the maximum distance;

[0235] This indicates the infimum operation, used to find the minimum distance.

[0236] 3) Calculate the total loss function

[0237] The total loss function is calculated by weighting and summing the cross-entropy loss, shape-aware loss, and boundary alignment loss terms according to preset weight coefficients. This integrates pixel-level classification error, shape regularity constraints, and boundary alignment constraints to optimize the model, thereby improving the accuracy and boundary consistency of dust pollution source segmentation. This is expressed as:

[0238]

[0239] The total loss function is represented by the combination of pixel-level classification error, shape regularity constraints, and boundary alignment constraints to improve the accuracy and boundary consistency of dust pollution source segmentation.

[0240] Cross-entropy loss is used to measure pixel-level classification error, based on the predicted segmentation mask. and the real segmentation mask The calculation measures the difference between the predicted probability and the true label for each pixel, and is expressed as... ;

[0241] Represents the true segmentation mask exist Intensity value at pixel location;

[0242] Represents the predicted segmentation mask exist Intensity value at pixel location;

[0243] The weight coefficient represents the shape-aware loss term and is used to control the impact of shape compactness on the total loss. The larger the value, the more the model tends to predict regular shapes. It is preferably set to 0.3.

[0244] This represents the weighting coefficient of the boundary alignment loss term, used to strengthen the consistency between the predicted boundary and the true boundary. The larger the value, the more the model pays attention to the boundary error. It is preferably set to 0.2.

[0245] In a specific implementation, the adaptive fusion enhancement of the dust pollution source area is as follows:

[0246] In image enhancement of dust pollution sources, existing methods often neglect the difference enhancement between pollution sources and backgrounds, resulting in insufficient detail in the pollution source area or excessive background enhancement in the enhanced image, affecting visual consistency and subsequent analysis. Conventional fusion methods directly apply global enhancement operators, which cannot adaptively adjust the enhancement strategies of pollution sources and backgrounds, and lack multi-scale detail enhancement and texture preservation mechanisms, making it difficult to highlight the pollution source while maintaining the naturalness of the background.

[0247] This invention achieves region adaptive fusion enhancement by fusing multi-scale features, segmentation masks, and attention mechanisms. The specific steps are as follows:

[0248] 1) Enhanced multi-scale details of pollution source areas

[0249] Multi-scale detail enhancement of the pollution source region is performed using encoder multi-scale feature maps and predicted segmentation masks. This is achieved by calculating and weighting multi-scale detail maps to enhance the visibility of details in the pollution source region, as shown below:

[0250]

[0251] In the formula, Represents multi-scale detail enhancement images exist The intensity value at the pixel location is used to enhance the visibility of details in the pollution source area. The pixel value normalized feature map of the multi-scale detail enhancement image is shown in Figure 9.

[0252] Indicates the first The scale-based detail enhancement weights adaptively adjust the contribution of features at different scales to detail enhancement, and are calculated as follows: ;

[0253] For the first The scale detail enhancement weights of the fully connected layer weight vectors are used to map global average pooling features to scalar scores and are trainable parameters.

[0254] For the first The scale detail enhancement weights of the fully connected layer bias term, used to adjust the score, are trainable parameters.

[0255] Indicates the encoder's first... Stage feature map The feature map obtained after upsampling is Intensity value at pixel location;

[0256] Indicates an upsampling operation. The term represents the encoder's first term. Stage feature map Perform an upsampling operation to make its spatial size... The upsampling method uses bilinear interpolation.

[0257] This represents the Gaussian kernel function used to smooth the feature map. The standard deviation of the Gaussian kernel is preferably set to 2 to control the scale of detail extraction.

[0258] This indicates a global average pooling operation. The encoder will be the first Stage feature map Spatial dimension pooling Preserve channel dimensions;

[0259] Represents the predicted segmentation mask exist The intensity value at the pixel location is used to focus on the pollution source area.

[0260] It should be noted that details are emphasized with weight. By calculating trainable parameters, the contribution of features at different scales is adaptively adjusted to highlight pollution source details. Gaussian kernel smoothing is used to extract high-frequency detail components in the feature map, and the enhanced region is multiplied by the segmentation mask to limit the enhanced region.

[0261] 2) Background lighting uniformity and texture injection

[0262] By employing illumination component estimation and texture synthesis strategies, the background area is homogenized and a realistic dust texture is injected to maintain a natural background, as shown below:

[0263]

[0264] In the formula, Represents background blending image exist The intensity value at the pixel location is used to maintain the naturalness of the background;

[0265] Represents the predicted segmentation mask exist The intensity value at a pixel location is used to distinguish between the background and the pollution source area;

[0266] Represents background illumination components exist The intensity value at a pixel location is calculated as follows: It is used to smooth background areas and suppress noise and uneven lighting.

[0267] To represent the Gaussian kernel function used to simulate background lighting, the standard deviation of the Gaussian kernel is preferably set to 20 to control the smoothness of the background.

[0268] This represents the texture blending intensity, with a preferred value of 0.2, which controls the extent of texture addition.

[0269] This represents the Laplacian operator, used to extract high-frequency details from textures;

[0270] Represents background texture blocks exist The intensity value at the pixel location is a texture block randomly selected from the dust texture library, which includes typical background textures.

[0271] In its implementation, the dust texture library contains a large number of cropped blocks of background areas from real dust pollution source images, such as... Pixels are obtained by randomly selecting a texture block and resizing it by scaling or interpolation. Size ensures that the texture blends seamlessly with the background.

[0272] It should be noted that the background lighting components Large-scale Gaussian convolution estimation smooths the background region to reduce noise; the Laplacian operator extracts high-frequency texture details, ensuring the injected texture blends naturally with the background. A segmentation mask is used to protect the contamination source region from modification.

[0273] 3) Adaptive weighted fusion output

[0274] Based on the spatial attention weight map and edge enhancement mask, adaptive weight fusion is performed on the detail enhancement image and the background fusion image to generate the final enhanced dust pollution source image, ensuring the prominence of the pollution source and visual consistency, as shown below:

[0275]

[0276] In the formula, This indicates an enhanced image of dust pollution sources. exist The intensity value at the pixel location outputs an enhanced image that highlights the pollution source, presents a natural background, and has high visual consistency.

[0277] Represents the enhanced fusion weight matrix exist The value at the pixel location is used to balance detail enhancement and background blending, and is calculated as follows: ;

[0278] The first fusion coefficient is preferably 0.6, which is used to control the influence of spatial attention weights on the fusion weight matrix. Higher values ​​emphasize the detail enhancement of pollution source-related areas, making the model more reliant on the attention mechanism to locate key areas.

[0279] The second fusion coefficient, preferably 0.4, is used to control the influence of the edge enhancement mask on the fusion weight matrix. Higher values ​​strengthen the fusion transition in the boundary region, ensuring natural fusion of the pollution source outline and reducing artifacts.

[0280] It should be noted that the enhanced fusion weight matrix By dynamically adjusting attention and edge masking, the image tends to enhance details in pollution source areas and blend with the background in background areas, ensuring visual consistency and ultimately enhancing the image to effectively highlight pollution source details while maintaining a natural background.

[0281] In a specific implementation, the model training is carried out as follows:

[0282] During the model training phase, the trainable parameters of the encoder and decoder modules are first initialized. The weights are initialized using a random normal distribution, and the optimizer is set to an adaptive moment estimation optimizer. The initial learning rate is 0.001, and the batch size is adjusted to 16 based on hardware resources.

[0283] In each iteration, the batch of synthesized enhanced dust pollution source images and their corresponding real segmentation masks are input. The segmentation mask is predicted by forward propagation, and the difference between the model output and the real mask is evaluated based on the total loss function combined with cross-entropy loss, shape-aware loss and boundary alignment loss.

[0284] During backpropagation, gradients are calculated and model parameters are updated using the chain rule to minimize the total loss function, while gradient pruning is employed to prevent gradient explosion.

[0285] Iterative training monitors performance on the validation set. After each training round, the loss and segmentation accuracy metrics on the validation set, such as intersection-over-union ratio and Dice coefficient, are calculated to evaluate the model's generalization ability.

[0286] The criteria for stopping iterations are based on the early stopping mechanism and the maximum number of iterations. Specifically, when the validation set loss does not decrease or decreases by less than the threshold of 0.0001 within 20 consecutive iterations, training is terminated early to avoid overfitting.

[0287] If the early stopping condition is not triggered, training will continue until the preset maximum number of iterations, 1000, is reached.

[0288] After training, the best-performing model parameters on the validation set are saved for image enhancement tasks to ensure that the model can effectively handle the multi-scale characteristics and blurred boundaries of dust pollution source images.

[0289] In a specific implementation, S5 is as follows:

[0290] After the model training is completed, the specific process for enhancing newly acquired dust pollution source images is as follows:

[0291] First, image deblurring and enhancement processing is performed on the newly input original image. Illumination components are estimated and reflection components are extracted through multi-scale Gaussian kernel convolution to generate an enhanced image of dust pollution sources.

[0292] Then, adaptive morphological operations based on local gradient magnitude are applied to obtain morphologically enhanced dust images to restore blurred boundary details and suppress noise;

[0293] Then, the morphologically enhanced dust image is input into the trained dust pollution source image enhancement model. The model uses an encoder to extract multi-scale features and a decoder to generate a predictive segmentation mask. The predictive segmentation mask identifies the dust pollution source region and the background region in the image.

[0294] Finally, a region-adaptive fusion enhancement operation is performed: first, multi-scale detail enhancement is applied to the pollution source region by upsampling the feature maps from each stage of the encoder and multiplying them with the predicted segmentation mask to highlight fine particles and contours; then, illumination homogenization is applied to the background region and realistic dust background texture is injected to maintain naturalness; finally, the enhanced detail image and the background fusion image are dynamically fused using a spatial attention weight map and an edge enhancement mask to generate a result such as... Figure 10 The final enhanced image of dust pollution sources shown effectively improves the visibility, boundary integrity, and texture realism of dust pollution sources in the new image, while ensuring a natural transition of the background area. It is suitable for real-time image enhancement tasks of environmental monitoring systems or mobile devices.

[0295] Example 2

[0296] A dust pollution source image enhancement processing system, comprising a dust pollution source image enhancement processing method, including:

[0297] Data acquisition module: Acquires images of dust pollution sources through camera equipment, records corresponding environmental parameters, annotates the acquired images, and forms a dust pollution source image dataset;

[0298] Data preprocessing module: includes contour enhancement unit and composite enhancement unit;

[0299] The system receives data from the dust pollution source image dataset in the data acquisition module, passes it sequentially through the contour enhancement unit and the synthesis enhancement unit, preprocesses the data in the dataset, and outputs the preprocessed data.

[0300] Dust pollution source image enhancement module: includes a dust pollution source image enhancement unit and an optimization unit;

[0301] The system receives preprocessed data from the data preprocessing module, then outputs enhanced dust pollution source images through the dust pollution source image enhancement unit. Finally, the optimization unit optimizes the output of the dust pollution source image enhancement unit.

[0302] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A method for image enhancement of a dust pollution source, characterized in that, The method comprises the following steps: S1, collecting dust pollution source images through on-site deployed camera equipment, recording corresponding environmental parameters, and labeling the images, the labeling categories including dust pollution source area and background area, forming a dust pollution source image dataset; S2, estimating the illumination component and reflection component of the dust pollution source images in the dataset through multi-scale Gaussian kernel convolution, then fusing the components of different scales to obtain a multi-scale enhanced image, and enhancing the irregular shape contour information of the multi-scale enhanced image through adaptive morphological operation to generate a morphological enhanced image; S3, performing elastic transformation on the morphological enhanced image through a generated random deformation field, and injecting real dust texture features by combining with a texture synthesis technology to obtain a synthetic enhanced image; S4, constructing a dust pollution source image enhancement model, inputting the synthetic enhanced image into the model, first extracting multi-scale features through a staged down-sampling encoder, outputting a fused multi-scale feature map, then enhancing the pollution source features and retaining boundary details through a decoder, outputting a jump connection fused feature, calculating a loss function according to the corresponding prediction segmentation mask output by the encoder and the labeled real mask to optimize the model, and then performing region adaptive fusion enhancement based on the fused multi-scale feature map, the prediction segmentation mask and the attention mechanism to output an enhanced dust pollution source image; training and optimizing the model to obtain a trained dust pollution source image enhancement model; S5, inputting a newly collected dust pollution source image into the trained model to generate a final enhanced dust pollution source image.

2. The method according to claim 1, wherein the method is characterized by, S1 is as follows: The camera equipment is a high-resolution equipment, including fixed monitoring cameras and mobile unmanned aerial vehicle platforms; The collected data covers multiple forms under different time periods, weather conditions and light environments, including diffusion, aggregation and drift, while recording corresponding environmental parameters; The images are finely labeled by combining manual labeling with semi-automatic tools, the dust pollution source area refers to the core area directly showing dust, smoke or particulate matter diffusion in the image, the background area refers to the non-pollutant part, including the sky, buildings and vegetation; during the labeling process, an environmental expert guides labelers to draw the pollution source boundary pixel by pixel using image labeling software to generate a binary segmentation mask, in which the pollution source area is marked as foreground and the background area is marked as background; The labeled image data is used to construct a dataset, and the labeled dataset is divided into a training set, a validation set and a test set.

3. The method according to claim 1, wherein the method is characterized by, S2 is as follows: The original dust pollution source image is subjected to illumination component estimation through multi-scale Gaussian kernel convolution, and the reflection component is calculated in the logarithmic domain, then the fusion weight of different scales is adaptively fused to output a multi-scale enhanced image; Then, the multi-scale enhanced image is subjected to morphological opening operation based on a local gradient amplitude adaptive structure element, the value of the structure element is dynamically adjusted according to the gradient threshold value to eliminate noise in the image and strengthen the irregular shape boundary to generate a morphological enhanced image.

4. The method according to claim 1, wherein the method is characterized by, S3 is as follows: A smooth random deformation field is generated based on a radial basis function interpolation, and a deformation offset is calculated by a random offset weight of a control point, an irregular shape change of a fugitive dust pollution source is simulated by applying a non-rigid transformation to a morphological enhancement image to generate a deformation enhancement image, and the shape diversity of the training sample is increased. A synthetic enhancement image is obtained by randomly selecting a texture block from a fugitive dust texture library, extracting high-frequency details of the texture based on a Laplacian operator, and seamlessly integrating the texture block into the deformation enhancement image according to a texture fusion strength.

5. The method according to claim 1, wherein the method is characterized by, The operations in the encoder are as follows: The encoder main body is a structure of phased serial downsampling, including L stages, each stage being composed of convolution, activation function and pooling layer; Each stage gradually extracts features through a series of convolution, activation function and pooling operations, the input of the first stage being the synthetic enhancement image, the output of the first stage being used as the input of the second stage, and the output of the previous stage being used as the input of the next stage; Then the output of each stage is processed through a plurality of parallel hollow convolution branches, the number of branches corresponding to the number of stages, each branch using a convolution kernel with different hollow rates to perform a hollow convolution operation on the output of the last stage of the encoder, and combining a rectified linear unit activation function to generate an output feature map; Finally, the output feature maps of the hollow convolution of each branch are upsampled to adjust the spatial size, and are weighted and fused according to adaptive fusion weights and attention masks to generate a fused multi-scale feature map.

6. The method according to claim 1, wherein the method is characterized by, The operations in the decoder are as follows: An attention weight map is generated using a spatial attention mechanism, the fused multi-scale feature map is deconvolved and upsampled, and an edge enhancement mask is combined to highlight the boundary area, dynamically weight the pollution source related area in the feature map, enhance the feature response of the pollution source area, and generate an attention weighted upsampled feature map; wherein the edge enhancement mask is generated based on the gradient amplitude of the multi-scale enhancement image; Then the output feature maps of different stages of the encoder are adjusted by convolution, and are connected to the attention weighted upsampled feature map by a jump connection, and are fused according to adaptive weights and edge enhancement attention maps, thereby retaining fine-grained boundary information, to generate a jump connection fusion feature map; wherein the edge enhancement attention map is generated based on the gradient amplitude of the multi-scale enhancement image.

7. The method according to claim 1, wherein the method is characterized by, A composite loss function is designed by combining a shape perception term, a boundary distance constraint and a cross-entropy loss, and the calculation process of the loss function is as follows: The skip connection fused feature map output by the decoder is first processed by a convolution and a sigmoid activation function to obtain a predicted segmentation mask. The skip connection fused feature map output by the decoder is first processed by a convolution and a sigmoid activation function to obtain a predicted segmentation mask. The compactness ratio is calculated based on the area and boundary perimeter of the predicted segmentation mask and the labeled real segmentation mask, and the empty mask is processed by an indicator function, and then the shape perception loss term is calculated; The Hausdorff distance between the predicted boundary and the real boundary is calculated, and the maximum value of the two-way distance is taken as the boundary alignment loss to constrain the consistency of the predicted boundary and the real boundary, and then the boundary alignment loss term is calculated; The difference between the predicted probability of each pixel point and the real label is calculated based on the predicted segmentation mask and the labeled real segmentation mask, and the cross-entropy loss measuring the pixel-level classification error is obtained; The cross-entropy loss, the shape perception loss term and the boundary alignment loss term are weighted and summed to calculate the total loss function.

8. The method according to claim 1, wherein the method is characterized by, Region adaptive The fusion enhancement process is as follows: The pollution source region is reinforced with multi-scale details by using the encoder multi-scale feature map and the predicted segmentation mask. The detail visibility of the pollution source region is enhanced by calculating a multi-scale detail map and performing weighted fusion, so as to obtain a multi-scale detail reinforced image. The background region is subjected to uniform illumination processing and injection of real dust texture by using the illumination component estimation and texture synthesis strategy, so as to generate a background fusion image. The multi-scale detail reinforced image and the background fusion image are subjected to adaptive weight fusion based on the spatial attention weight map and the edge enhancement mask, so as to generate a final enhanced dust pollution source image.

9. The method according to claim 1, wherein the method is characterized by, The model training process is as follows: The data in the data set is divided into a training set, a validation set and a test set. The data in the training set and the validation set are used to iteratively train the model. Firstly, the trainable parameters of the encoder module and the decoder module are initialized. The weights are initialized by using a random normal distribution. The optimizer, the initial value of the learning rate and the batch size are set. In each iteration, the batch of synthesized enhanced images and their corresponding real segmentation masks are input. The predicted segmentation mask is calculated by forward propagation. The difference between the model output and the real segmentation mask is evaluated based on the loss function. During the back propagation process, the gradient is calculated and the model parameters are updated by using the chain rule to minimize the loss function. Meanwhile, the gradient clipping strategy is used to prevent gradient explosion. The iterative training monitors the performance on the validation set. The loss and segmentation accuracy on the validation set are calculated after each training round to evaluate the generalization ability of the model. The early stopping mechanism and the maximum number of iterations are used to set the judgment conditions for stopping iteration. After the training is completed, the model parameters with the best performance on the validation set are saved to obtain the trained model. The trained model is tested by using the data in the test set.

10. A dust pollution source image enhancement processing system, which performs the dust pollution source image enhancement processing method according to any one of claims 1-9, characterized in that: The data acquisition module acquires dust pollution source images by using a camera device and records corresponding environmental parameters. The acquired images are labeled to form a dust pollution source image data set. The data preprocessing module includes a contour enhancement unit and a synthetic enhancement unit. The data preprocessing module receives the data set of dust pollution source images from the data acquisition module, and sequentially passes through the contour enhancement unit and the synthetic enhancement unit to preprocess the data in the data set and output the preprocessed data. The dust pollution source image enhancement module includes a dust pollution source image enhancement unit and an optimization unit. The dust pollution source image enhancement module receives the preprocessed data output by the data preprocessing module, passes through the dust pollution source image enhancement unit to output an enhanced dust pollution source image, and passes through the optimization unit to optimize the output of the dust pollution source image enhancement unit.

Images