Coal mining face dust mist image defogging method and system based on latent diffusion model

Abstract

A dust and fog image defogging method and system for a coal mining face based on a latent diffusion model, the method comprising compressing a dust and fog image to a latent space through a self-encoder and adding a smooth loss to optimize the latent space features; directly adding the intermediate features of the encoder to the decoder through residual connection and feature fusion to enrich the features of the processed image; using a residual phase block to extract the structural features of the source image and performing multi-scale feature cross fusion, together with the time step, to embed a denoising U-Net for guided denoising; combining the perception loss and the SSIM loss and adding a contrast learning loss to optimize the result. The system comprises a camera, a sample collector and a defogging trainer. The method and system can effectively defog the coal mining face image, provide a high-definition image source for subsequent target detection and recognition, and improve the level of coal mine safety hazard identification and monitoring.

Description

Technical Field

[0001] This invention relates to a method and system for defogging images of coal mining faces based on a latent diffusion model, belonging to the field of image defogging technology. Background Technology

[0002] Due to the large amount of coal dust and water mist generated during coal mining, the working environment in mines is often in a state of high dust and fog, especially in the area where the coal mining machine is cutting coal. The dust and fog are so dense that the image quality captured by the camera equipment is often greatly affected. The coal seam features in the images are blurred and details are lost, which greatly limits the identification of coal and rock and the precise operation of intelligent coal mining machines.

[0003] Traditional image dehazing algorithms are mostly based on image enhancement, image restoration, and image fusion methods, such as histogram equalization, homomorphic filtering, and dark channel prior algorithms. However, these methods are often inadequate for handling images in complex mining environments and cannot meet real-world needs. In recent years, deep learning has been widely applied in image processing, and more and more image dehazing solutions are adopting deep learning-based approaches. Deep learning-based image dehazing algorithms learn the mapping relationship from foggy to clear images by training a large number of foggy and clear image pairs. This greatly optimizes the performance of image dehazing algorithms and achieves relatively good dehazing results on natural datasets. However, current mainstream deep learning dehazing algorithms cannot consider the complex fog conditions in mines, and their dehazing results on mine images are not ideal. For example, diffusion models have powerful image generation capabilities, and controlling the denoising network of diffusion models can endow them with good image restoration capabilities. However, existing diffusion models have long training times, insufficient constraints on the denoising network, and the generated images are random, making them unsuitable for direct application to image dehazing tasks in high-dust and foggy mining environments. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for defogging images of coal mining faces based on a latent diffusion model. This method and system can effectively defog images of coal mining faces, providing high-definition image sources for subsequent target detection and recognition, and improving the level of coal mine safety hazard identification and monitoring.

[0005] To achieve the above objectives, the present invention provides a method for defogging dust images of coal mining faces based on a latent diffusion model, comprising the following steps:

[0006] S1. The dust and fog image is compressed into the latent space using an autoencoder, and a smoothing loss is added to optimize the latent space features;

[0007] S2. By using residual connection and feature fusion, intermediate features of the encoder are directly added to the decoder to enrich the features of the processed image.

[0008] S3. Use residual phase blocks to extract structural features from the source image and perform multi-scale feature cross-fusion, then embed them into a denoising U-Net along with the time step for guided denoising;

[0009] S4. Combine the perceptual loss and SSIM loss, and add the contrastive learning loss to optimize the results.

[0010] Furthermore, the autoencoder in S1 adopts a U-Net structure design. The first three layers are used to extract image features, and each layer consists of two residual blocks and downsampling. The fourth layer is used to compress image features and consists of a single residual block. The residual block consists of two 3×3 convolutional layers and a residual connection.

[0011] The smoothing loss, through progressively varying regularization, forces the ratio between changes in any input latent space and changes in the output image to remain constant during the diffusion training step. The objective of the smoothing loss is expressed by the following formula:

[0012]

[0013] The objective of smoothing loss is a fixed-size change Δε over random noise ε, i.e., x T The change Δx T lead to Changes on It is stable, with a maximum value of constant C, and ε ~ N(0,I); This is the final result of back diffusion; x T This is the result of the forward diffusion at time step T; to incorporate it into the loss function, the above equation needs to be transformed as follows:

[0014]

[0015] The transformed formula applies to a certain time step t in the diffusion process. A ratio is formed between Δε and Δε, and its ratio is... Proportional, of which α t =1-β t ,β t The diffusion rate is preset for the time step t; based on the purpose of the smoothing loss, the smoothing loss function is constructed using the following regularization loss:

[0016]

[0017] in, The normalized pixel intensity to a unit length pixel is the sampled pixel intensity. The variation in ε is the random noise of normal sampling, and a is the exponential moving average calculated online.

[0018] Furthermore, the encoder in S2 consists of a single residual block in the first layer for feature expansion, and each of the last three layers consists of two residual blocks and downsampling for feature decoding.

[0019] The intermediate features f1, f2, and f3 generated by the first three layers of the encoder are fused through a connection layer, a 3×3 convolutional layer, batch normalization (BN), and ReLU activation function, and then directly input into the decoder. The specific process is as follows:

[0020] S2.1 Perform channel-level concatenation on the first two features f1 and f2: F1 = Concat(f1, f2), where Concat(·) is the concatenation operation;

[0021] S2.2 Perform a 3×3 convolution operation on the obtained F1: F2 = Conv(F1), where Conv(·) is a 3×3 convolution operation;

[0022] S2.3. Perform batch normalization and ReLU activation function on the obtained F2:

[0023] F3 = BN(F2);

[0024] F4 = ReLU(F3);

[0025] Where BN(·) is batch normalization and ReLU(·) is the ReLU activation function; F3 and F4 are used as inputs for the second fusion, and the fusion operations from S2.1 to S2.3 are repeated with the intermediate feature value f3 generated by the third layer of the encoder to obtain the final output.

[0026] Furthermore, the residual phase block in S3 extracts feature information from the spatial and frequency domains. The spatial domain feature extraction consists of two 3×3 convolutions, a ReLU activation function, and a residual connection, yielding spatial domain features f. spa The image is separated into amplitude information (Amp) and phase information (Phase) using Fast Fourier Transform (FFT). The phase information is further extracted using 1×1 convolution and ReLU activation function. Then, the processed phase and amplitude information are mapped back to the image space using Inverse Fast Fourier Transform (IFFT) to obtain the frequency information (f). fre Finally, f spa and f fre Connect the elements and generate output features through a 1×1 convolution; specifically:

[0027] Spatial domain feature extraction is represented as:

[0028] f spa =ReLU(Conv 3×3 (ReLU(Conv 3×3 (fin ))))+Conv 1×1 (f in );

[0029] Among them, Conv 3×3 (·) represents a 3×3 convolution operation, Conv 1×1 (·) represents a 1×1 convolution operation, and ReLU(·) represents the ReLU activation function;

[0030] Frequency domain feature extraction is represented as:

[0031] f fre =IFFT(Conv 1×1 (ReLU(Conv 1×1 (f Phase ))),f Amp );

[0032] Among them, f Phase For phase spectrum information, f Amp For amplitude spectrum information, IFFT(·) is the inverse fast Fourier transform;

[0033] The output is generated as follows:

[0034] f out =Conv 1×1 (Concat(f spa ,f fre )).

[0035] To generate conditional guidance information suitable for each layer of the denoising U-Net, multiple residual phase block and downsampling operations are performed on the source image. The resulting information at different scales is cross-fused to generate multi-scale fused features. The generated structural features can serve as guiding conditions for diffusion model denoising. Specifically, each scale feature is first extracted through a 3×3 convolution, then simple channel concatenation is performed with other scale features, and finally fused through a 1×1 convolution to achieve a combination of high-resolution and low-resolution features.

[0036] Furthermore, the specific process of S4 is as follows:

[0037] S4.1, Perceptual loss is represented as: L per_loss =L percep (I pred ,I gt )+L percep (I xrec ,I gt ); where L percep For the perceptual loss function, I pred For the image sampled in the current iteration, I xrecI is a reference image used for encoding and decoding a real image for reconstruction. gt It is a real image;

[0038] S4.2, SSIM loss is expressed as: L SSIM =2-SSIM(I pred ,I gt )-SSIM(I xrec ,I gt ); where SSIM(·) is the SSIM calculation function;

[0039] S4.3. Combine the SSIM loss and the perception loss into a content loss, expressed as: L Content =λ1L SSIM +λ2L per_loss Where λ1 and λ2 are the weights of the SSIM loss and the perception loss, respectively;

[0040] S4.4. Through course contrastive regularization, negative samples are divided into easy negative, hard negative, and extremely hard negative samples, and different weights are assigned for regularization. The course contrastive learning loss is expressed as:

[0041]

[0042] Among them, E i =||V i (I)-V i (f(I,θ))||1, f(·,θ) represents a dehazing network with parameter θ, V i (·), i = 1, 2, ..., n are the i-th hidden features extracted from pre-trained VGG-19, the number of negative samples is k, {ζ i Let} be the hyperparameter set, J be the positive sample image, and S be the hyperparameter set. q For images that are not prone to negative samples, The weights of samples that are not easily negative at time t0; The definition is as follows:

[0043]

[0044] Where γ is a hyperparameter, the weight of difficult negative samples is set to 1+γ, the weight of extremely difficult negative samples is set to 1-γ, and the weight of easy negative samples is set to the number of negative samples k; in contrastive learning, extremely difficult negative samples will be dynamically optimized into difficult negative samples.

[0045] S4.5 The overall loss function is expressed as: L = λ Content L Content +λ reg L reg +λ CR L CR ;

[0046] Where, λ Content , λ reg , λ CR These are the weights for content loss, smoothing loss, and course-contrast learning loss, respectively.

[0047] The present invention also provides a defogging system for coal mining face dust and fog images based on a potential diffusion model, including a camera, a sample collector, and a defogging trainer;

[0048] The camera and its memory card are used to acquire and store dust and fog image sample data of the coal mining face for training the defogging model. They are installed at different locations on the coal mining face to acquire images containing dust and fog.

[0049] The sample acquisition device is used to receive sample data acquired by the camera, perform filtering and conventional image preprocessing, and send it to the dehazing trainer for training.

[0050] The dehazing trainer includes a potential diffusion model module and a performance optimization module, which are used to dehaze the collected samples and optimize the model.

[0051] This invention improves the speed of the diffusion model by compressing images into the latent space using an autoencoder and incorporates a smoothing loss to smooth the features of the latent space to optimize the model. Then, intermediate features from the encoder are fused through convolution and directly added to the decoder to enrich the features of the processed image. Conditional constraint features are obtained by extracting residual phase blocks of structural features from the source image. Finally, the model is optimized by combining perceptual loss, SSIM loss, and contrastive learning loss. Clear images are used as positive samples, and blurred images are categorized into easy negative, hard negative, and extremely hard negative samples based on contrast difficulty. Contrastive learning improves model performance by moving the processed image closer to positive samples and away from negative samples. This invention effectively dehazes images of coal mining faces, providing a high-definition image source for subsequent target detection and recognition, and improving the level of coal mine safety hazard identification and monitoring. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the workflow of the method of the present invention;

[0053] Figure 2 (a) is a schematic diagram of the encoder and decoder of the present invention, and (b) is a schematic diagram of the residual block structure of the present invention;

[0054] Figure 3 This is a schematic diagram of the residual phase block of the present invention;

[0055] Figure 4 This is a schematic diagram of the multi-scale feature fusion of the present invention;

[0056] Figure 5 This is a schematic diagram of the course comparison learning regularization of the present invention;

[0057] Figure 6 (a) is a fogged image, and (b) is a defogging image processed by the present invention. Detailed Implementation

[0058] The invention will now be further described with reference to the accompanying drawings.

[0059] like Figure 1 As shown, a method for defogging dust images in coal mining faces based on a latent diffusion model includes the following steps:

[0060] S1. The dust and fog image is compressed into the latent space using an autoencoder, and a smoothing loss is added to optimize the latent space features;

[0061] S2. By using residual connection and feature fusion, intermediate features of the encoder are directly added to the decoder to enrich the features of the processed image.

[0062] S3. Use residual phase blocks to extract structural features from the source image and perform multi-scale feature cross-fusion, then embed them into a denoising U-Net along with the time step for guided denoising;

[0063] S4. Combine the perceptual loss and SSIM loss, and add the contrastive learning loss to optimize the results.

[0064] As a preferred implementation method, such as Figure 2 As shown in (a), the autoencoder in S1 adopts a U-Net structure design. The first three layers are used to extract image features, each layer consisting of two residual blocks and downsampling. The fourth layer is used to compress image features and consists of a single residual block. Figure 2 As shown in (b), the residual block consists of two 3×3 convolutional layers and a residual connection;

[0065] The smoothing loss, through progressively varying regularization, forces the ratio between changes in any input latent space and changes in the output image to remain constant during the diffusion training step. The objective of the smoothing loss is expressed by the following formula:

[0066]

[0067] The objective of smoothing loss is a fixed-size change Δε over random noise ε, i.e., x T The change Δx T lead to Changes on It is stable, with a maximum value of constant C, and ε ~ N(0,I); This is the final result of back diffusion; x TThis is the result of the forward diffusion at time step T; to incorporate it into the loss function, the above equation needs to be transformed as follows:

[0068]

[0069] The transformed formula applies to a certain time step t in the diffusion process. A ratio is formed between Δε and Δε, and its ratio is... Proportional, of which α t =1-β t ,β t The diffusion rate is preset for the time step t; based on the purpose of the smoothing loss, the smoothing loss function is constructed using the following regularization loss:

[0070]

[0071] in, The normalized pixel intensity to a unit length pixel is the sampled pixel intensity. The variation in ε is the random noise of normal sampling, and a is the exponential moving average calculated online.

[0072] like Figure 2 As shown in (a), the first layer of the encoder in S2 consists of a single residual block for feature expansion, and each of the last three layers consists of two residual blocks and downsampling for feature decoding;

[0073] The intermediate features f1, f2, and f3 generated by the first three layers of the encoder are fused through a connection layer, a 3×3 convolutional layer, batch normalization (BN), and ReLU activation function, and then directly input into the decoder. The specific process is as follows:

[0074] S2.1 Perform channel-level concatenation on the first two features f1 and f2: F1 = Concat(f1, f2), where Concat(·) is the concatenation operation;

[0075] S2.2 Perform a 3×3 convolution operation on the obtained F1: F2 = Conv(F1), where Conv(·) is a 3×3 convolution operation;

[0076] S2.3. Perform batch normalization and ReLU activation function on the obtained F2:

[0077] F3 = BN(F2);

[0078] F4 = ReLU(F3);

[0079] Where BN(·) is batch normalization and ReLU(·) is the ReLU activation function; F3 and F4 are used as inputs for the second fusion, and the fusion operations from S2.1 to S2.3 are repeated with the intermediate feature value f3 generated by the third layer of the encoder to obtain the final output.

[0080] As a preferred implementation method, such as Figure 3 As shown, the residual phase block in S3 extracts feature information from the spatial and frequency domains. The spatial domain feature extraction consists of two 3×3 convolutions, a ReLU activation function, and a residual connection, yielding spatial domain features f. spa The image is separated into amplitude information (Amp) and phase information (Phase) using Fast Fourier Transform (FFT). The phase information is further extracted using 1×1 convolution and ReLU activation function. Then, the processed phase and amplitude information are mapped back to the image space using Inverse Fast Fourier Transform (IFFT) to obtain the frequency information (f). fre Finally, f spa and f fre Connect the elements and generate output features through a 1×1 convolution; specifically:

[0081] Spatial domain feature extraction is represented as:

[0082] f spa =ReLU(Conv 3×3 (ReLU(Conv 3×3 (f in ))))+Conv 1×1 (f in );

[0083] Among them, Conv 3×3 (·) represents a 3×3 convolution operation, Conv 1×1 (·) represents a 1×1 convolution operation, and ReLU(·) represents the ReLU activation function;

[0084] Frequency domain feature extraction is represented as:

[0085] f fre =IFFT(Conv 1×1 (ReLU(Conv 1×1 (f Phase ))),f Amp );

[0086] Among them, f Phase For phase spectrum information, f Amp For amplitude spectrum information, IFFT(·) is the inverse fast Fourier transform;

[0087] The output is generated as follows:

[0088] fout =Conv 1×1 (Concat(f spa ,f fre )).

[0089] To generate conditional guiding information suitable for each layer of the denoising U-Net, multiple residual phase block and downsampling operations are performed on the source image. The resulting information at different scales is cross-fused to generate multi-scale fused features. The generated structural features can serve as guiding conditions for diffusion model denoising. Specifically, the cross-fusion is as follows: Figure 4 As shown, each scale feature is first extracted through a 3×3 convolution, then connected to other scale features through simple channel convolution, and finally fused through a 1×1 convolution to achieve a combination of high-resolution and low-resolution features.

[0090] Furthermore, the specific process of S4 is as follows:

[0091] S4.1, Perceptual loss is represented as: L per_loss =L percep (I pred ,I gt )+L percep (I xrec ,I gt ); where L percep For the perceptual loss function, I pred For the image sampled in the current iteration, I xrec I is a reference image used for encoding and decoding a real image for reconstruction. gt It is a real image;

[0092] S4.2, SSIM loss is expressed as: L SSIM =2-SSIM(I pred ,I gt )-SSIM(I xrec ,I gt ); where SSIM(·) is the SSIM calculation function;

[0093] S4.3. Combine the SSIM loss and the perception loss into a content loss, expressed as: L Content =λ1L SSIM +λ2L per_loss Wherein, λ1 and λ2 are the weights of SSIM loss and perception loss, respectively, and in this embodiment they are set to 0.5 and 1, respectively.

[0094] S4.4. Contrastive learning improves network performance by moving closer to positive samples and further away from negative samples. The contrastive learning strategy minimizes the L1 distance between the anchor point and positive samples while maximizing the L1 distance between the anchor point and negative samples to achieve regularization. However, ordinary contrastive learning cannot cope with complex mining environments. Curriculum-based contrastive regularization, as a preferred implementation method, categorizes negative samples into easy negative, hard negative, and extremely hard negative samples, assigning different weights for regularization. Figure 5 As shown, the contrastive learning loss of the course is expressed as:

[0095]

[0096] Among them, E i =||V i (I)-V i (f(I,θ))||1, f(·,θ) represents a dehazing network with parameter θ, V i (·), i = 1, 2, ..., n are the i-th hidden features extracted from pre-trained VGG-19, the number of negative samples is k, {ζ i Let} be the hyperparameter set, J be the positive sample image, and S be the hyperparameter set. q For images that are not prone to negative samples, The weights of samples that are not easily negative at time t0; The definition is as follows:

[0097]

[0098] Where γ is a hyperparameter, the weight of difficult negative samples is set to 1+γ, the weight of extremely difficult negative samples is set to 1-γ, and the weight of easy negative samples is set to the number of negative samples k; in contrastive learning, extremely difficult negative samples will be dynamically optimized into difficult negative samples.

[0099] S4.5 The overall loss function is expressed as: L = λ Content L Content +λ reg L reg +λ CR L CR ;

[0100] Where, λ Content , λ reg , λ CR These are the weights for content loss, smoothing loss, and course-contrast learning loss, respectively.

[0101] The present invention also provides a defogging system for coal mining face dust and fog images based on a potential diffusion model, including a camera, a sample collector, and a defogging trainer;

[0102] The camera and its memory card are used to acquire and store dust and fog image sample data of the coal mining face for training the defogging model. They are installed at different locations on the coal mining face to acquire images containing dust and fog.

[0103] The sample acquisition device is used to receive sample data acquired by the camera, perform filtering and conventional image preprocessing, and send it to the dehazing trainer for training.

[0104] The dehazing trainer includes a potential diffusion model module and a performance optimization module, which are used to dehaze the collected samples and optimize the model.

[0105] The samples are fed into the latent diffusion model module, trained using the method provided in this invention, and then optimized using the performance optimization module, which includes the loss function optimization method designed in this invention. The latent diffusion model module and the performance optimization module interact to complete the construction of a method and system for defogging coal face dust and fog images based on the latent diffusion model.

[0106] To verify the performance of the present invention, such as Figure 6 As shown in (a), the dehazing capability of the algorithm of this invention was tested on the dataset, and the results are as follows. Figure 6 As shown in (b). From Figure 6 As can be seen from the comparison between (a) and (b), the present invention achieves a significant dehazing effect on foggy images.

Claims

1. A method for defogging dust images in coal mining faces based on a latent diffusion model, characterized in that, Includes the following steps: S1. The dust and fog image is compressed into the latent space using an autoencoder, and a smoothing loss is added to optimize the latent space features; S2. By using residual connection and feature fusion, intermediate features of the encoder are directly added to the decoder to enrich the features of the processed image. S3. Use residual phase blocks to extract structural features from the source image and perform multi-scale feature cross-fusion, then embed them into a denoising U-Net along with the time step for guided denoising; S4. Combine the perceptual loss and SSIM loss, and add the contrastive learning loss to optimize the results; The residual phase block in S3 extracts feature information from the spatial domain and the frequency domain, wherein the spatial domain feature extraction consists of two... The spatial domain features are obtained by using convolution, ReLU activation function and residual connection. The amplitude information is separated by Fast Fourier Transform (FFT). and phase information Phase information is further obtained through Feature extraction is performed using convolution and ReLU activation functions. Then, the processed phase and amplitude information is mapped back to the image space using Inverse Fast Fourier Transform (IFFT) to obtain frequency information. Finally and Connect, and through a Convolution generates output features; specifically: Spatial domain feature extraction is represented as: ； in, for Convolution operation, for Convolution operation, It is the ReLU activation function; Frequency domain feature extraction is represented as: ； in, For phase spectrum information, For amplitude spectrum information, This is the inverse fast Fourier transform; The output is generated as follows: 。 2. The method for defogging coal face dust and fog images based on a latent diffusion model according to claim 1, characterized in that, The autoencoder in S1 adopts a U-Net structure design. The first three layers are used to extract image features, and each layer consists of two residual blocks and downsampling. The fourth layer is used to compress image features and consists of a single residual block. The residual block consists of two... It consists of a convolutional layer and a residual connection; The smoothing loss, through progressively varying regularization, forces the ratio between changes in any input latent space and changes in the output image to remain constant during the diffusion training step. The objective of the smoothing loss is expressed by the following formula: ； The target of smoothing loss is random noise. Fixed size change ,Right now Changes on lead to Changes on It is stable, and its highest value is a constant. , ; This is the final result of reverse diffusion; This is the result of the forward diffusion at time step T; to incorporate it into the loss function, the above equation needs to be transformed as follows: ； The transformed formula applies to a certain time step in the diffusion process. , and A ratio is formed between them, and their ratio is... Proportional, of which , , For time step The preset diffusion rate; based on the purpose of the smoothing loss, the smoothing loss function is constructed using the following regularization loss: ； in, The normalized pixel intensity to a unit length pixel is the sampled pixel intensity. Changes on This is random noise from normal sampling. This is an exponential moving average calculated online.

3. The method for defogging coal face dust and fog images based on a latent diffusion model according to claim 2, characterized in that, The encoder in S2 consists of a single residual block in the first layer for feature expansion, and each of the last three layers consists of two residual blocks and a downsampling unit for feature decoding. Intermediate features generated by the first three layers of the encoder , , Through the connection layer, The convolutional layer, batch normalized normalization (BN), and ReLU activation function are fused and directly input into the decoder. The specific process is as follows: S2.1, the first two features , Perform channel-level splicing: ,in, For splicing operations; S2.2, the obtained conduct Convolution operation: ,in, for Convolution operation; S2.3, the obtained Batch normalization and ReLU activation function are performed separately: ； ； in, For batch normalization, The ReLU activation function is used; F3 and F4 are used as inputs for the second fusion, along with intermediate feature values ​​generated by the third layer of the encoder. Repeat steps S2.1 to S2.3 to obtain the final output.

4. The method for defogging coal face dust and fog images based on a latent diffusion model according to claim 3, characterized in that, The specific process of S4 is as follows: S4.1, Perceived loss is expressed as: ;in, For the perceptual loss function, For the image sampled in the current iteration, This is a reference image used for encoding and then decoding a real image for reconstruction. It is a real image; S4.2, SSIM loss is expressed as: ;in, The function is used to compute SSIM; S4.

3. Combine the SSIM loss and the perception loss into a content loss, expressed as: ;in, and These are the weights for the SSIM loss and the perception loss, respectively. S4.

4. Through course contrastive regularization, negative samples are divided into easy negative, hard negative, and extremely hard negative samples, and different weights are assigned for regularization. The course contrastive learning loss is expressed as: ； in, , The parameter is Defogging network, , Extracted from pre-trained VGG-19 There are hidden features, and the number of negative samples is . , For hyperparameter set, For positive sample images, For images that are not prone to negative samples, For the first The weights of samples that are not easily negative at all times; The definition is as follows: ； in, As a hyperparameter, the weights of difficult negative samples are set to... Extremely difficult negative samples are set The weight of easily negative samples is set to the number of negative samples. In contrastive learning, extremely difficult negative samples are dynamically optimized into difficult negative samples. S4.5 The overall loss function is expressed as: ; in, , , These are the weights for content loss, smoothing loss, and course-contrast learning loss, respectively.

5. A system for defogging images of coal mining faces based on a latent diffusion model as described in any one of claims 1 to 4, comprising a camera, characterized in that, It also includes a sample collector and a defogging trainer; The camera and its memory card are used to acquire and store dust and fog image sample data of the coal mining face for training the defogging model. They are installed at different locations on the coal mining face to acquire images containing dust and fog. The sample acquisition device is used to receive sample data acquired by the camera, perform conventional image preprocessing such as filtering, and send it to the dehazing trainer for training. The dehazing trainer includes a potential diffusion model module and a performance optimization module, which are used to dehaze the collected samples and optimize the model.

Images