A method and system for image dehazing and enhancement in dusty and foggy environments at fully mechanized coal mining faces.

Abstract

A method and system for image dehazing enhancement in a dusty and foggy environment at a fully mechanized coal mining face is disclosed. The method includes acquiring a foggy image to be processed, encoding the foggy image using an encoder to obtain coded features, inputting the obtained coded features into a bottleneck layer integrated by a DM module to achieve feature coupling in the spatial and frequency domains, obtaining fused features, and decoding the fused features using a decoder to output the target dehazed image. The system includes a camera and a dual-domain coupled image dehazing network model module. This invention enhances the representation of uneven fog features and global feature perception through deformable convolution and amplitude-phase guided strategies, improving the clarity of image feature extraction, restoring image details and structure, and improving the model's dehazing effect and generalization ability.

Description

Technical Field

[0001] This invention relates to a method and system for image defogging and enhancement in a dusty and foggy environment at a fully mechanized coal mining face, belonging to the field of image processing technology. Background Technology

[0002] Image dehazing is a key task in computer vision, aiming to restore clear, high-quality images degraded by fog and haze. In the dusty environment of fully mechanized coal mining faces, the entire process—from coal mining, loading, transportation, support, goaf treatment, and roadway excavation—is mechanized, generating large amounts of coal dust and particulate matter. Furthermore, water spraying and misting are commonly used to suppress gas and dust concentrations, further exacerbating the fog and dust buildup. Traditional image dehazing algorithms typically rely on hand-designed image priors. These methods obtain prior knowledge by statistically analyzing the relationship between blurred and clear images, but their specific assumptions limit their generalization ability when dealing with complex blurred images and unseen scenes.

[0003] Significant progress has been made in image dehazing with advancements in deep convolutional neural networks. These methods, trained under supervision on large datasets, have achieved excellent results in peak signal-to-noise ratio and structural similarity index on SOTS-Indoor. Recently, the introduction of Transformers into image dehazing has enhanced the ability to effectively capture global information and long-range dependencies in images, improving feature representation and achieving state-of-the-art performance. Nevertheless, the quadratic complexity of self-attention remains a significant computational challenge. Furthermore, texture features may exhibit strong correlations or extensions in specific directions, and the degradation caused by haze is almost entirely reflected in the amplitude component, while the phase component difference between hazy and clear images is small.

[0004] In summary, the main problems with existing image dehazing techniques are: fixed-size convolutions are difficult to achieve inter-channel interaction and cannot effectively extract multi-scale information, thus making it difficult to accurately capture the distribution of haze and reconstruct clear texture details in dense haze areas; they mainly focus on extracting spatial features in the spatial domain, resulting in poor image feature extraction performance. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for image defogging and enhancement in a dusty and foggy environment at a fully mechanized coal mining face. This method and system can improve the clarity of image feature extraction, restore the details and structure of the image, enhance the generalization ability of the model, and ensure safe mining in coal mines.

[0006] To achieve the above objectives, the present invention provides a method for image dehazing and enhancement in a dusty and foggy environment at a fully mechanized coal mining face, comprising the following steps:

[0007] S1. Obtain the foggy image to be processed, and encode the foggy image using an encoder to obtain the encoded features from the foggy image;

[0008] S2. Input the obtained coding features into the bottleneck layer integrated by the DM module to realize the coupling of coding features in the spatial domain and frequency domain, and obtain the fused features;

[0009] S3. The obtained fused features are decoded by the decoder to output the target dehazed image.

[0010] Furthermore, the specific process of S1 is as follows:

[0011] S1.1. Using the RESIDE dataset, obtain foggy images and clear images respectively. Randomly crop the foggy images to 256×256 and horizontally flip the foggy images with a 50% probability to train the defogging network.

[0012] S1.2 The dehazing network consists of an encoder, a bottleneck layer, and a decoder. The encoder performs three downsampling operations, doubling the feature channels and halving the width and height. Each downsampling operation consists of four ResBlock blocks, each composed of two 3×3 convolutions and an intermediate GELU activation function repeated four times. The encoder progressively extracts the encoded features corresponding to the hazy image features through the ResBlock blocks. Each ResBlock block has a stronger ability to extract encoded features than the previous one. The multi-layered structure of the encoder can capture and process image degradation features at different scales, thereby improving the robustness of dehazing.

[0013] Furthermore, the DM module in S2 includes a convolutional layer, a DCM module, an APGM module, and an SRU module; the feature fusion process is as follows:

[0014] S2.1. The deformable convolution operator (DCM) effectively aggregates spatial context information and reconstructs texture details in heavily hazy areas, for the input features. The DC-h and DC-w operations are applied respectively to extract context information from the height and width directions, as shown in the following formulas:

[0015]

[0016] In the formula, ω i The parameter o represents the deformable convolution. c For the width and height dimensions in channel c i Learnable offset on;

[0017] S2.2 Calculate the average value of all pixels in each channel, double the number of channels by applying a 1×1 point convolution, and perform a softmax operation to obtain the weight ω:

[0018]

[0019] ω = softmax(conv(X) avg ));

[0020] In the formula, X avg X represents DC-h With X DC-w For the average value of all pixels in each channel, conv represents a 1×1 convolution, and ω represents X. DC-h With X DC-w Weight array;

[0021] S2.3, Relate the weights ω to X respectively. DC-h and X DC-w Multiply, then convolve to produce the final output:

[0022] X DCM =ω[0]*X DC-h +ω[1]*X DC-w ;

[0023] In the formula, ω[0] and ω[1] represent X respectively. DC-h and X DC-h The weights, X DCM This is the output of the DCM module; the DCM module effectively enhances the model's ability to capture the distribution of haze, which is beneficial for extracting clear feature information.

[0024] S2.4. The amplitude and phase guidance module (APGM) is used to guide the extraction of global features by utilizing prior knowledge of amplitude and phase.

[0025] S2.4.1 For the input feature map The input features are transformed from the spatial domain to the frequency domain using the Fast Fourier Transform (FFT), yielding the amplitude and phase information as follows:

[0026]

[0027] phase = atan(B, A);

[0028] In the formula, For a Fast Fourier Transform, A and B represent the real and imaginary parts, respectively, and atan() is the arctangent function;

[0029] S2.4.2. Low-frequency filtering is performed on the amplitude, and high-frequency filtering is performed on the phase. This is achieved by applying a residual block (ResBlock) consisting of two convolutional layers, along with an intermediate LeakyReLU activation function, to the amplitude and phase components. Subsequently, the original and modulated amplitude and phase information are combined using an inverse Fourier transform to generate a feature map containing both phase and amplitude modulation.

[0030] amb′=conv 1*1 (LeakyReLu(conv 1*1 (amb)));

[0031] phase′=conv 1*1 (LeakyReLu(conv 1*1 (phase)));

[0032]

[0033] In the formula, e i This represents a complex exponent. It is the inverse operation of the Fast Fourier Transform;

[0034] S2.4.3. Connect the modulated feature maps along the channel dimension, then perform a 1×1 point convolution to reduce the number of channels by half, obtaining the final output:

[0035] X APGM =conv 1*1 (concat(X amb ,X angel ));

[0036] In the formula, X APGM This indicates the output of the APGM module; this operation enables mutual guidance between amplitude and phase, enhancing the model's ability to capture global information more effectively.

[0037] S2.5. Fusion of spatial and frequency domain features: The input features are processed in parallel through a 1×1 point convolution, a DCM module, and an APGM module. Then, the output features are combined using addition to generate reconstructed information that combines spatial and frequency domain features.

[0038] S2.5.1. The output features obtained are combined through addition operations to generate reconstructed information combining the spatial and frequency domains, represented as:

[0039] F = conv 1×1 (X)+X DCM (X)+X APGM (X);

[0040] In the formula, conv 1×1X represents the output feature of a convolution operation with a 1×1 kernel. DCM and X APGM This indicates the output characteristics of the DCM module and the APGM module;

[0041] S2.5.2. Spatial domain and frequency domain coupling is achieved through the Spatial Reconstruction Module (SRU) for input features. First, normalization is performed using BatchNorm. Then, the Sigmoid function is used to map the weight map to the range (0,1). Next, a soft thresholding operation is performed, setting weights above a certain threshold to 1 to obtain information weights, and weights below the threshold to 0 to form non-information weights. These weights are multiplied by the input feature F to generate two weighted features: an information-rich feature and a less information-rich feature, as shown below:

[0042] F w =σ(BatchNorm(F));

[0043]

[0044] In the formula, σ is the sigmoid function, F w F represents the weights of the input feature F, where F1 and F2 represent information-rich features and information-poor features, respectively.

[0045] S2.5.3 To further refine the information content and optimize the space, a reconstruction operation is performed to fuse the two weighted features and enhance the information flow between them. Finally, the reconstructed features are connected to generate the final feature map. Represented as:

[0046]

[0047] Further, the method of S3 is as follows: the decoder performs three upsampling operations, reducing the feature channels to half their original size, and doubling both the width and height; each upsampling operation consists of four ResBlock blocks, each ResBlock block consisting of two 3×3 convolutions and an intermediate GELU activation function repeated four times; after the last ResBlock block, a 3×3 convolution maps the features to the RGB space to output a dehazed image; the specific process includes:

[0048] S3.1, Denote the fusion features as The encoder is denoted as Decoder, and the image is obtained after processing by the decoder. Represented as:

[0049]

[0050] S3.2 To ensure consistency between the dehazed and clear images at the pixel level and in the frequency domain, L1 loss, frequency domain L1 loss, amplitude contrast loss, and phase contrast loss are used as the main optimization objectives. The loss functions are as follows:

[0051]

[0052] In the formula, Y represents the dehazed image, and Y represents the clear image. Let A() and P() represent L1 loss, frequency domain L1 loss, amplitude contrast loss, and phase contrast loss, respectively. Let A() and P() represent amplitude and phase information, respectively. The overall network loss is expressed as follows:

[0053]

[0054] In the formula, λ1, λ2, and λ3 are set to 0.1, 0.9, and 0.2 respectively based on experience to balance the dual-domain training.

[0055] An image defogging enhancement system for a coal mine fully mechanized mining face under dust and fog conditions includes a camera and a module based on a dual-domain coupled image defogging network model.

[0056] The camera is used to capture images of dust and fog environment in coal mine fully mechanized mining faces, save them to a local path, and write the paths of foggy images to a database for image defogging.

[0057] The dual-domain coupled image dehazing network model module is used to dehaze foggy images to obtain restored images; by iteratively training the dual-domain coupled image dehazing network on the RESIDE-Outdoor dataset, an optimized dehazing model is obtained.

[0058] The dual-domain coupled image dehazing network is trained as follows: 30 iterations are performed on an RTX 3090 GPU with a training batch size of 8. To improve model convergence, the Adam optimizer is used, with parameters β1 and β2 set to 0.9 and 0.99, respectively. To enhance training efficiency and performance, a cosine annealing decay strategy is employed to reduce the learning rate from 2×10⁻⁶. -4 The number of iterations was gradually reduced to 1×10. -6 ;

[0059] The trained network model is loaded, the path of the foggy image is obtained from the database, the foggy image is read, and the defogging image is output after one forward propagation, thus completing the construction of the image defogging enhancement system for the dust and fog environment of the coal mine fully mechanized mining face.

[0060] This invention enhances the representation and global feature perception of uneven haze features through deformable convolution and amplitude-phase guided strategies. The deformable convolution operator DCM, based on a channel-interactive spatial non-uniform strategy, captures features along the image's height and width channels, effectively aggregating spatial context information to process uneven haze distribution and reconstruct fine texture details in heavily hazy regions. The amplitude-phase guided module APGM performs low-pass filtering on the amplitude component and high-pass filtering on the phase component to effectively extract global features. By combining the DCM and APGM modules, the DM module is inserted into the encoder-decoder bottleneck layer to overcome the obstacles in achieving spatial and frequency domain fusion. Finally, supervised training of the encoder-decoder architecture is performed to output a dehazed image. This invention improves the model's clarity in image feature extraction, restores image details and structure, enhances the model's dehazing effect and generalization ability, and ensures safe coal mining. Attached Figure Description

[0061] Figure 1 This is a flowchart of the process of the method of the present invention;

[0062] Figure 2 This is a diagram of the DM module of the present invention;

[0063] Figure 3 This is a diagram of the DCM module of the present invention;

[0064] Figure 4 This is a diagram of the APGM module of the present invention;

[0065] Figure 5 This is a diagram of the SRU module of the present invention;

[0066] Figure 6 This is a schematic diagram illustrating the working principle of the system of the present invention;

[0067] Figure 7 This is a comparison chart of the dehazing effects of the embodiments of the present invention and existing algorithms on the SOTS-Outdoor dataset;

[0068] Figure 8 These are comparison images of the defogging effect of embodiments of the present invention in a fully mechanized coal mining face environment. Detailed Implementation

[0069] The invention will now be further described with reference to the accompanying drawings.

[0070] like Figure 1 As shown, a method for image dehazing and enhancement in a dusty and foggy environment at a fully mechanized coal mining face includes the following steps:

[0071] S1. Obtain the foggy image to be processed, and encode the foggy image using an encoder to obtain the encoded features from the foggy image;

[0072] S2. Input the obtained coding features into the bottleneck layer integrated by the DM module to realize the coupling of coding features in the spatial domain and frequency domain, and obtain the fused features;

[0073] S3. The obtained fused features are decoded by the decoder to output the target dehazed image.

[0074] As a preferred embodiment, the specific process of S1 is as follows:

[0075] S1.1. Using the RESIDE dataset, obtain foggy images and clear images respectively. Randomly crop the foggy images to 256×256 and horizontally flip the foggy images with a 50% probability to train the defogging network.

[0076] S1.2 The dehazing network consists of an encoder, a bottleneck layer, and a decoder. The encoder performs three downsampling operations, doubling the feature channels and halving the width and height. Each downsampling operation consists of four ResBlock blocks, each composed of two 3×3 convolutions and an intermediate GELU activation function repeated four times. The encoder progressively extracts the encoded features corresponding to the hazy image features through the ResBlock blocks. Each ResBlock block has a stronger ability to extract encoded features than the previous one. The multi-layered structure of the encoder can capture and process image degradation features at different scales, thereby improving the robustness of dehazing.

[0077] Furthermore, such as Figure 2 As shown, the DM module in S2 includes a convolutional layer, a DCM module, an APGM module, and an SRU module; the feature fusion process is as follows:

[0078] S2.1, such as Figure 3 As shown, the deformable convolution operator (DCM) effectively aggregates spatial context information and reconstructs texture details in heavily hazy regions, for the input features. The DC-h and DC-w operations are applied respectively to extract context information from height and width, as shown in the following formulas:

[0079]

[0080] In the formula, ω i The parameter o represents the deformable convolution. c For the width and height dimensions in channel c i Learnable offset on;

[0081] S2.2 Calculate the average value of all pixels in each channel, double the number of channels by applying a 1×1 point convolution, and perform a softmax operation to obtain the weight ω:

[0082]

[0083] ω = softmax(conv(X) avg ));

[0084] In the formula, X avg X represents DC-h With X DC-w For the average value of all pixels in each channel, conv represents a 1×1 convolution, and ω represents X. DC-h With X DC-w Weight array;

[0085] S2.3, Relate the weights ω to X respectively. DC-h and X DC-w Multiply, then convolve to produce the final output:

[0086] X DCM =ω[0]*X DC-h +ω[1]*X DC-w ;

[0087] In the formula, ω[0] and ω[1] represent X respectively. DC-h and X DC-h The weights, X DCM This is the output of the DCM module; the DCM module effectively enhances the model's ability to capture the distribution of haze, which is beneficial for extracting clear feature information.

[0088] S2.4, such as Figure 4 As shown, the amplitude and phase guided module (APGM) utilizes prior knowledge of amplitude and phase to guide the extraction of global features:

[0089] S2.4.1 For the input feature map The input features are transformed from the spatial domain to the frequency domain using the Fast Fourier Transform (FFT), yielding the amplitude and phase information as follows:

[0090]

[0091] phase = atan(B,A);

[0092] In the formula, For a Fast Fourier Transform, A and B represent the real and imaginary parts, respectively, and atan() is the arctangent function;

[0093] S2.4.2. Low-frequency filtering is performed on the amplitude, and high-frequency filtering is performed on the phase. This is achieved by applying a residual block (ResBlock) consisting of two convolutional layers, along with an intermediate LeakyReLU activation function, to the amplitude and phase components. Subsequently, the original and modulated amplitude and phase information are combined using an inverse Fourier transform to generate a feature map containing both phase and amplitude modulation.

[0094] amb′=conv 1*1 (LeakyReLu(conv 1*1 (amb)));

[0095] phase′=conv 1*1 (LeakyReLu(conv 1*1 (phase)));

[0096]

[0097] In the formula, e i This represents a complex exponent. It is the inverse operation of the Fast Fourier Transform;

[0098] S2.4.3. Connect the modulated feature maps along the channel dimension, then perform a 1×1 point convolution to reduce the number of channels by half, obtaining the final output:

[0099] X APGM =conv 1*1 (concat(X amb ,X angel ));

[0100] In the formula, X APGM This indicates the output of the APGM module; this operation enables mutual guidance between amplitude and phase, enhancing the model's ability to capture global information more effectively.

[0101] S2.5. Fusion of spatial and frequency domain features: The input features are processed in parallel through a 1×1 point convolution, a DCM module, and an APGM module. Then, the output features are combined using addition to generate reconstructed information that combines spatial and frequency domain features.

[0102] S2.5.1. The output features obtained are combined through addition operations to generate reconstructed information combining the spatial and frequency domains, represented as:

[0103] F = DW 1×1 (X)+X DCM (X)+X APGM (X);

[0104] In the formula, conv 1×1X represents the output feature of a convolution operation with a 1×1 kernel. DCM and X APGM This indicates the output characteristics of the DCM module and the APGM module;

[0105] S2.5.2, such as Figure 5 As shown, spatial and frequency domain coupling is achieved through the spatial reconstruction module SRU for the input features. First, normalization is performed using BatchNorm. Then, the Sigmoid function is used to map the weight map to the range (0,1). Next, a soft thresholding operation is performed, setting weights above a certain threshold to 1 to obtain information weights, and weights below the threshold to 0 to form non-information weights. These weights are multiplied by the input feature F to generate two weighted features: an information-rich feature and a less information-rich feature, as shown below:

[0106] F w =σ(BatchNorm(F));

[0107]

[0108] In the formula, σ is the sigmoid function, F w F represents the weights of the input feature F, where F1 and F2 represent information-rich features and information-poor features, respectively.

[0109] S2.5.3 To further refine the information content and optimize the space, a reconstruction operation is performed to fuse the two weighted features and enhance the information flow between them. Finally, the reconstructed features are connected to generate the final feature map. Represented as:

[0110]

[0111] Further, the method of S3 is as follows: the decoder performs three upsampling operations, reducing the feature channels to half their original size, and doubling both the width and height; each upsampling operation consists of four ResBlock blocks, each ResBlock block consisting of two 3×3 convolutions and an intermediate GELU activation function repeated four times; after the last ResBlock block, a 3×3 convolution maps the features to the RGB space to output a dehazed image; the specific process includes:

[0112] S3.1, Denote the fusion features as The encoder is denoted as Decoder, and the image is obtained after processing by the decoder. Represented as:

[0113]

[0114] S3.2 To ensure consistency between the dehazed and clear images at the pixel level and in the frequency domain, L1 loss, frequency domain L1 loss, amplitude contrast loss, and phase contrast loss are used as the main optimization objectives. The loss functions are as follows:

[0115]

[0116] In the formula, Y represents the dehazed image, and Y represents the clear image. Let A() and P() represent L1 loss, frequency domain L1 loss, amplitude contrast loss, and phase contrast loss, respectively. Let A() and P() represent amplitude and phase information, respectively. The overall network loss is expressed as follows:

[0117]

[0118] In the formula, λ1, λ2, and λ3 are set to 0.1, 0.9, and 0.2 respectively based on experience to balance the dual-domain training.

[0119] like Figure 6 As shown, the present invention provides an image defogging enhancement system for coal mine fully mechanized mining faces under dust and fog conditions, including a camera and a module based on a dual-domain coupled image defogging network model;

[0120] The camera is used to capture images of dust and fog environment in coal mine fully mechanized mining faces, save them to a local path, and write the paths of foggy images to a database for image defogging.

[0121] The dual-domain coupled image dehazing network model module is used to dehaze foggy images to obtain restored images; by iteratively training the dual-domain coupled image dehazing network on the RESIDE-Outdoor dataset, an optimized dehazing model is obtained.

[0122] The dual-domain coupled image dehazing network is trained as follows: 30 iterations are performed on an RTX 3090 GPU with a training batch size of 8. To improve model convergence, the Adam optimizer is used, with parameters β1 and β2 set to 0.9 and 0.99, respectively. To enhance training efficiency and performance, a cosine annealing decay strategy is employed to reduce the learning rate from 2×10⁻⁶. -4 The number of iterations was gradually reduced to 1×10. -6 ;

[0123] The trained network model is loaded, the path of the foggy image is obtained from the database, the foggy image is read, and the defogging image is output after one forward propagation, thus completing the construction of the image defogging enhancement system for the dust and fog environment of the coal mine fully mechanized mining face.

[0124] Example: A comparison of the dehazing effect of the method of the present invention with existing dehazing algorithms on the SOTS-Outdoor dataset. The results are as follows. Figure 7 As shown in (a) to (f). Among them, Figure 7 (a) is the original image to be processed. Figure 7 (b) to (d) are images processed by the methods proposed by Dehame, OKNet and DeaNet-CR, respectively. It can be seen that gray mottled artifacts and haze still exist in the red rectangle. Figure 7 (e) is the image processed by the present invention. Figure 7 (f) represents a clear image. The images processed above are compared with... Figure 7 As can be seen from the comparison in (f), the method of the present invention generates the most natural image while retaining more background details, without causing color deviation and edge blurring, and the reconstructed image is closer to the real image.

[0125] To verify the generalization ability of this method, this invention acquired foggy images of the coal mine fully mechanized mining face environment, such as... Figure 8 As shown in (a), the image is input into the system of this invention for processing to obtain a fog-free image, as follows. Figure 8 As shown in (b), the image processed by this invention exhibits significantly improved visual quality. Specifically, the method of this invention can effectively remove smog caused by factors such as water vapor spraying and successfully restore the detailed texture information in the image, thereby significantly improving the image clarity and overall visual quality.

Claims

1. A method for image dehazing and enhancement in a dusty and foggy environment at a fully mechanized coal mining face, characterized in that, Includes the following steps: S1. Obtain the foggy image to be processed, and encode the foggy image using an encoder to obtain the encoded features from the foggy image; S2. Input the obtained coding features into the bottleneck layer integrated by the DM module to realize the coupling of coding features in the spatial domain and frequency domain, and obtain the fused features; S3. The obtained fused features are decoded by the decoder to output the target dehazed image; The DM module in S2 includes a convolutional layer, a DCM module, an APGM module, and an SRU module; The fusion process of the fusion features is as follows: S2.

1. The deformable convolution operator (DCM) effectively aggregates spatial context information and reconstructs texture details in heavily hazy areas, for the input features. The DC-h and DC-w operations are applied respectively to extract context information from the height and width directions, as shown in the following formulas: In the formula, ω i The parameter o represents the deformable convolution. c For the width and height dimensions in channel c i Learnable offset on; S2.2 Calculate the average value of all pixels in each channel, double the number of channels by applying a 1×1 point convolution, and perform a softmax operation to obtain the weight ω: ω=softmax(conv(X avg )); In the formula, X avg X represents DC-h With X DC-w For the average value of all pixels in each channel, conv represents a 1×1 convolution, and ω represents X. DC-h With X DC-w Weight array; S2.3, Relate the weights ω to X respectively. DC-h and X DC-w Multiply, then convolve to produce the final output: X DCM =ω[0]*X DC-h +ω[1]*X DC-w ; In the formula, ω[0] and ω[1] represent X respectively. DC-h and X DC-h The weights, X DCM This is the output of the DCM module; S2.

4. The amplitude and phase guidance module (APGM) is used to guide the extraction of global features by utilizing prior knowledge of amplitude and phase. S2.4.1 For the input feature map The input features are transformed from the spatial domain to the frequency domain using the Fast Fourier Transform (FFT), yielding the amplitude and phase information as follows: phase = atan(B,A); In the formula, For a Fast Fourier Transform, A and B represent the real and imaginary parts, respectively, and atan() is the arctangent function; S2.4.

2. Low-frequency filtering is performed on the amplitude, and high-frequency filtering is performed on the phase. This is achieved by applying a residual block (ResBlock) consisting of two convolutional layers, along with an intermediate LeakyReLU activation function, to the amplitude and phase components. Subsequently, the original and modulated amplitude and phase information are combined using an inverse Fourier transform to generate a feature map containing both phase and amplitude modulation. with′=conv 1*1 (LeakyReLu(conv 1*1 (with))); phase′=conv 1*1 (LeakyReLu(conv 1*1 (phase))); In the formula, e i This represents a complex exponent. It is the inverse operation of the Fast Fourier Transform; S2.4.

3. Connect the modulated feature maps along the channel dimension, then perform a 1×1 point convolution to reduce the number of channels by half, obtaining the final output: X APGM =conv 1*1 (concat(X amb ,X angel )); In the formula, X APGM This indicates the output of the APGM module; S2.5, Fusion of spatial and frequency domain features: S2.5.

1. The output features obtained are combined through addition operations to generate reconstructed information combining the spatial and frequency domains, represented as: F=conv 1×1 (X)+X DCM (X)+X APGM (X); In the formula, conv 1×1 X represents the output feature of a convolution operation with a 1×1 kernel. DCM and X APGM This indicates the output characteristics of the DCM module and the APGM module; S2.5.

2. Spatial domain and frequency domain coupling is achieved through the Spatial Reconstruction Module (SRU) for input features. First, normalization is performed using BatchNorm. Then, the Sigmoid function is used to map the weight map to the range (0,1). Next, a soft thresholding operation is performed, setting weights above a certain threshold to 1 to obtain information weights, and weights below the threshold to 0 to form non-information weights. These weights are multiplied by the input feature F to generate two weighted features: an information-rich feature and a less information-rich feature, as shown below: F w =σ(BatchNorm(F)); In the formula, σ is the sigmoid function, F w F represents the weights of the input feature F, where F1 and F2 represent information-rich features and information-poor features, respectively. S2.5.3 Perform a reconstruction operation to connect the reconstructed features to generate the final feature map. Represented as:

2. The image dehazing and enhancement method for fully mechanized coal mining faces under dust and fog conditions according to claim 1, characterized in that, The specific process of S1 is as follows: S1.

1. Using the RESIDE dataset, obtain foggy images and clear images respectively. Randomly crop the foggy images to 256×256 and horizontally flip the foggy images with a 50% probability to train the defogging network. S1.2 The dehazing network consists of an encoder, a bottleneck layer, and a decoder. The encoder performs three downsampling operations, doubling the feature channels and halving the width and height. Each downsampling operation consists of four ResBlock blocks, each consisting of two 3×3 convolutions and an intermediate GELU activation function repeated four times. The encoded features corresponding to the foggy image features are gradually extracted through the ResBlock blocks of the encoder.

3. The image dehazing and enhancement method for fully mechanized coal mining faces under dust and fog conditions according to claim 1, characterized in that, The S3 method is as follows: the decoder performs three upsampling operations, reducing the feature channels to half their original size, and doubling both the width and height; each upsampling operation consists of four ResBlock blocks, each ResBlock consisting of two 3×3 convolutions and an intermediate GELU activation function repeated four times; after the last ResBlock, a 3×3 convolution maps the features to the RGB space to output a dehazed image; the specific process includes: S3.1, Denote the fusion features as The encoder is denoted as Decoder, and the image is obtained after processing by the decoder. Represented as: S3.2 To ensure consistency between the dehazed and clear images at the pixel level and in the frequency domain, L1 loss, frequency domain L1 loss, amplitude contrast loss, and phase contrast loss are used as the main optimization objectives. The loss functions are as follows: In the formula, Y represents the dehazed image, and Y represents the clear image. Let A() and P() represent L1 loss, frequency domain L1 loss, amplitude contrast loss, and phase contrast loss, respectively. Let A() and P() represent amplitude and phase information, respectively. The overall network loss is expressed as follows: In the formula, λ1, λ2, and λ3 are set to 0.1, 0.9, and 0.2 respectively based on experience to balance the dual-domain training.

4. A system for image defogging and enhancement in a fully mechanized coal mining face under dust and fog conditions as described in any one of claims 1 to 3, comprising a camera, characterized in that, It also includes a dual-domain coupled image dehazing network model module; The camera is used to capture images of dust and fog environment in coal mine fully mechanized mining faces, save them to a local path, and write the paths of foggy images to a database for image defogging. The dual-domain coupled image dehazing network model module is used to dehaze foggy images to obtain restored images; by iteratively training the dual-domain coupled image dehazing network on the RESIDE-Outdoor dataset, an optimized dehazing model is obtained.

Images