Depth guidance image defogging method and system for deep ground medical rescue

Abstract

The invention discloses a depth guidance image defogging method and system for deep ground medical rescue, and aims to improve the visibility of medical rescue monitoring images and the accuracy and real-time performance of structural information recovery in deep ground environments such as mines. By introducing a scale collaborative attention module, collaborative learning of multi-scale features is realized, scale confusion is relieved, and the cross-scale feature collaboration capability is enhanced; and meanwhile, an error perception guide module is adopted to fuse local feature difference and global dependence, and mutual promotion optimization is performed on defogging and depth estimation tasks, so that more accurate image restoration is realized in a region with a complicated structure or non-uniform fog distribution. According to the method, close collaborative optimization and performance improvement between image defogging and depth estimation tasks can be realized, and the problems of scale confusion and limited scale collaborative capability in multi-scale feature learning can be solved to a certain extent.

Description

Technical Field

[0001] This invention relates to an image dehazing method and system, specifically a depth-guided image dehazing method and system for intelligent perception and risk prediction of personnel behavior, physiological parameters and environmental quality in special environments such as mines and tunnels, applicable to deep-earth medical rescue, belonging to the field of artificial intelligence and safety monitoring technology. Background Technology

[0002] In deep-earth medical rescue scenarios, rescuers work for extended periods in complex environments such as underground tunnels, collapsed areas, or abandoned mines, generating a large amount of airborne particulate matter, including dust, particulate matter, and fine water droplets. These fine particles continuously disperse in the air, and the rescue environment often features dust suppression sprays or high humidity, further forming high concentrations of suspended fog, resulting in a large number of visible particles and fog droplets in the air. Simultaneously, the complex and variable lighting conditions in underground environments, including rescue lighting, LED flashlights, reflective walls, and reflections from metal equipment, combine to severely degrade the image quality captured by imaging equipment. Specifically, this manifests in the following ways: 1. Low contrast and overall darkening: Dust and fog droplets in the air scatter and absorb light, reducing the overall brightness of the image, causing uneven grayscale distribution, blurring of target edges, and loss of detail.

[0003] 2. Local overexposure and halo effect: Rescue lights, reflective metals and other strong light sources are prone to halo effect in foggy environments, resulting in excessive brightness in local areas, accompanied by glare and blurring, which increases the difficulty of image processing.

[0004] 3. Localized dense fog residue and reduced visibility: Fog droplets in sprayed or high-humidity areas are larger, resulting in enhanced light scattering and significant localized dense fog residue, which leads to target obstruction, blurred edges, and loss of image structural information.

[0005] In recent years, single-image dehazing has aimed to restore image sharpness and visual quality, but existing methods based on multi-scale feature extraction, feature fusion, and deep guided collaborative dehazing still have significant limitations. Regarding feature extraction, current methods typically employ parallel multi-scale convolution combined with attention mechanisms. However, features at different scales within the same spatial location often contain heterogeneous information, lacking effective cross-scale interaction and limiting the utilization of feature complementarity. In terms of feature fusion, features are usually fused through simple concatenation, weighting, or CNN-Transformer hybrid modules, which may introduce scale confusion and constrain scale collaboration. Traditional deep guided collaborative dehazing methods mainly rely on local difference mechanisms, enhancing local collaboration by focusing on regions with inconsistent local structures. However, they often ignore global information and areas with poor dehazing, resulting in limited performance in non-uniform fog environments. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a depth-guided image dehazing method and system for deep-earth medical rescue. It can achieve close collaborative optimization and performance improvement between image dehazing and depth estimation tasks, and can solve the problems of scale confusion and limited scale coordination ability in multi-scale feature learning to a certain extent. This will improve the accuracy of monitoring camera systems and intelligent detection systems in mines and enhance the level of coal mine safety production.

[0007] To achieve the above objectives, this depth-guided image dehazing method for deep-earth medical rescue specifically includes the following steps: S1. The system acquires foggy images in mines or deep underground environments using an industrial camera, and simultaneously uses a depth sensor to obtain optional depth information, providing input data for subsequent defogging and depth estimation. S2. Add a scale-coordinated attention module to the u-net network architecture, and introduce an error-aware guidance module in the downstream tasks. S3. In the upstream task, a cross-scale tensor coding module is first introduced to independently slice and store the features of different branches; then a multi-scale encoding module is introduced to enable the network to learn the interaction and complementarity of features of different scales in the same spatial location; finally, a multi-scale interleaving module is used to form a scale co-attention module by combining gating weighting with coordinate attention. S4. Introduce an error-aware guidance module in downstream tasks to enhance multimodal collaborative defogging capabilities through global guidance and local error awareness; S5. The system performs post-processing and optimization on the results after defogging and depth estimation, and displays the clear image and corresponding depth information on the display terminal. At the same time, the result data is stored in the database or local storage module for subsequent safety monitoring, personnel identification and auxiliary rescue.

[0008] Furthermore, the u-net network architecture in S2 includes three encoders and two decoders. Each encoder and decoder includes a ResBlock block. A scale-coordinated attention module is added to the residual of each ResBlock block, and only the first attention module is used to receive the depth information of the feature map.

[0009] Furthermore, the scale-coordinated attention module in S2 includes a cross-scale tensor coding module and a multi-scale coding module.

[0010] Furthermore, the computation steps of the S3 cross-scale tensor coding module are as follows: The input is represented as ,in This represents the input to the cross-scale tensor coding module. Indicates the scale quantity. Indicates batch size, Indicates the number of channels. Indicates the height of the feature map. Indicates the width of the feature map; First, features at different scales are extracted through three parallel branches: Local feature branches: In the formula: This represents the extracted local features; Represents the linear rectification activation function; express Convolution operation; Represents a linear transformation; Regional feature branches: In the formula: Represents contextual features; This represents a tensor that upsamples the image to the original image size, representing the contextual features. This indicates a size that is upsampled to twice the original size; Global feature branch: In the formula: Represents global features; This represents a tensor that upsamples the image to the original image size, representing global features. This indicates a size upsampled four times to the original size; Secondly, a new scale dimension D is introduced, and the extracted local, contextual, and global features are first assigned to their respective independent slices along the D dimension, as shown below: Finally, the individual slices are stitched together along the D dimension to form a cross-scale tensor, as shown below: In the formula: Represents a multi-scale tensor; This indicates that three features at different scales are stored independently on the new scale dimension D; Indicates the size of the newly added dimension; Each scale feature is preserved in an independent slice, with D=1 corresponding to the local feature slice, D=2 corresponding to the context feature slice, and D=3 corresponding to the global feature slice.

[0011] Furthermore, the calculation steps of the S3 multi-scale compilation module are as follows: Design a multi-scale 3D dilated convolution, setting the kernel size to 3×3×3, 3×5×5, and 3×7×7, and the dilation rate to 2. Apply these convolutions in parallel across the multi-scale tensors. The output of the multi-scale tensor encoding module is represented as follows: The output of the cross-scale tensor coding module is applied with three parallel 3D dilated convolutions of tensors with different receptive fields, as shown below: In the formula: , , These represent the three branches of parallel multi-scale 3D convolution; The symbols represent 3D convolution, with subscripts 3×3×3, 3×5×5, and 3×7×7 indicating the size of the 3D convolution kernel. ; This represents the dilation rate of a 3D convolution; The final output of the multi-scale compilation module is represented as follows: In the formula: This represents the final output of the three branches of the parallel multi-scale 3D convolution; This means reducing the dimensionality of the image in the D dimension to the original image size using average pooling. This represents the final output of the scale compilation module.

[0012] Furthermore, the error-aware guidance module of S4 emphasizes regions with inconsistent or difficult-to-handle structures from a global perspective through the global feature guidance module, and integrates the local differences captured by the local difference module through a gating fusion mechanism to achieve feature optimization.

[0013] Furthermore, the specific working process of the S4 error-aware guidance module is as follows: ① Global Feature Guidance Module: The global feature guidance module takes the features of the dehazed image or dehazed depth map and the corresponding real haze-free image or real haze-free depth map as input, as shown below: In the formula: This represents the dehazed image output by the model; Represents a true, fog-free image; The dehazed features are projected onto the query Q through a 1×1 convolution, while the true features are projected onto the key K and value V through independent 1×1 convolutions, as shown below: In the formula: This represents the calculation result of cross-attention; , and These represent the query, key, and value generated from the dehazed graph and the true haze-free graph, respectively. This represents multiple attention heads that divide the feature map; This indicates an operation that splices together the attention of multiple parties; Represents a nonlinear activation function; and These represent two different linear layers; Presentation layer normalization operation; This represents the normalized output; Indicates average pooling; express Transpose of; Indicates the normalization factor; Represents the normalization function; Then, a 3×3 convolution is applied, followed by a softmax normalization function to generate the final weight map. , means as follows: In the formula: This represents the generated weighted graph; Indicates a feedforward fully connected layer; For weighted graph Reversing the order is represented as follows: In the formula: Represents the error-aware weights of the dehazed image; Similarly, by inputting the dehazed depth map and the corresponding true fog-free depth map, the error-aware weights for depth estimation are obtained. ; Let the dehazed image be The real, fog-free image is The image containing fog is , , and The corresponding depth maps are as follows: and The error-aware weights generated by the global feature guidance module for the dehazed image and depth estimation are denoted as follows: = , The collaborative dehazing guided by the global feature guidance module is then represented as: In the formula: This represents the loss function of the depth estimation network guided by the global feature-driven module; This represents the loss function of the dehazing network guided by the global feature module; ②Local difference module: set up This indicates the weight of areas with poor defogging performance. The weights for depth estimation in areas with poor dehazing performance are represented by the following steps in the local difference module: In the formula: This represents a multilayer perceptron; Represents the normalization function; This represents a 3×3 two-dimensional convolution; The collaborative dehazing guided by the local difference module is then represented as: In the formula: and These represent the loss functions of the dehazing network and the depth estimation network guided by the local difference module, respectively. ③ Collaborative dehazing provides global guidance through a global feature guidance module and uses a local difference module to detect local inconsistencies for modeling. The weights of the two modules are gated and weighted to form the final error-aware guidance module: Let the weight of the area with poor final defogging effect be . , The weights for depth estimation in areas with poor dehazing performance are represented as follows. The final performance of dehazing using depth information assisted by the error-aware guidance module is expressed as follows: In the formula: and These represent the loss functions of the dehazing network and the depth estimation network guided by the error-aware guidance module, respectively.

[0014] A depth-guided image dehazing system for deep-earth medical rescue, based on a depth-guided image dehazing method for deep-earth medical rescue, includes an image acquisition unit, a dehazing processing unit, a depth estimation unit, and a post-processing and storage module; The image acquisition unit includes an industrial camera for acquiring foggy images and a depth sensor for simultaneously acquiring corresponding depth information; The dehazing unit includes a scale-coordinated attention module, which in turn includes a cross-scale tensor coding module and a multi-scale encoding module. The depth estimation unit includes an error-aware guidance module, which in turn includes a global feature guidance module and a local difference module. The post-processing and storage module includes a monitoring terminal.

[0015] Compared with existing technologies, this depth-guided image dehazing method for deep-earth medical rescue has the following advantages: 1. Employing a scale-cooperative attention mechanism to achieve multi-scale feature interaction and optimization: Unlike existing multi-scale feature extraction methods based on parallel convolution and attention, this invention achieves intrinsic interaction of cross-scale features through a multi-scale interleaving module, three-dimensional dilated convolution, and cross-scale tensor encoding. Combined with coordinate attention for spatial context modeling, this can alleviate the heterogeneity and confusion between features of different scales, and enhance the complementarity and cooperation of multi-scale features, thereby improving the completeness and accuracy of dehazing feature representation.

[0016] 2. An error-aware guidance module is used to achieve bidirectional mutual promotion between depth estimation and dehazing tasks: Unlike existing depth guidance methods that are based solely on local differences, this invention combines local differences with global guidance. Under the framework of mutual promotion between the two tasks, depth estimation and dehazing tasks guide and reinforce each other. This avoids the problem of local mechanisms ignoring global information and areas with unsatisfactory dehazing, and also improves the image restoration effect under non-uniform fog conditions.

[0017] 3. This invention combines a scale-coordinated attention mechanism with an error-aware guidance module, which outperforms existing methods in terms of feature utilization efficiency, spatial structure modeling capability, and task collaborative optimization, thereby achieving higher dehazing accuracy, stronger robustness, and better generalization ability. Attached Figure Description

[0018] Figure 1 This is a flowchart of the depth-guided image dehazing method for deep-earth medical rescue according to the present invention; Figure 2 This is an architecture diagram of the cross-scale tensor coding mechanism of this invention; Figure 3 This is an architectural diagram of the scale-coordinated attention mechanism of this invention; Figure 4 This is an architectural diagram of the coordinate attention frame mechanism of the present invention; Figure 5 This is an architectural diagram of the multi-scale weaving module of the present invention; Figure 6 This is an architecture diagram of the error-aware guidance module of the present invention; Figure 7 This is a comparison diagram of the algorithm of this invention with existing algorithms; Figure 8 This is a schematic diagram of the depth-guided image defogging system for deep-earth medical rescue according to the present invention. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings.

[0020] like Figure 1 As shown, this method for dehazing depth-guided images for deep-earth medical rescue specifically includes the following steps: S1. The system acquires foggy images in mines or deep-earth environments using an industrial camera, and simultaneously uses a depth sensor to obtain optional depth information, providing input data for subsequent defogging and depth estimation.

[0021] A schematic diagram of the depth-guided image dehazing system for deep-earth medical rescue of this invention is shown below. Figure 8As shown, the system acquires foggy images in deep-earth or mining environments using an industrial camera, and can optionally be equipped with a depth sensor (such as a binocular camera, structured light, ToF, or LiDAR) to simultaneously acquire corresponding depth information. The acquired data is transmitted to the subsequent processing unit via a high-speed interface. This module can operate stably under complex conditions such as low light, strong light interference, and dust dispersion, ensuring the integrity and reliability of the input data. The industrial camera provides high-resolution image information, while the depth sensor supplements spatial structure information. Both support timestamp alignment and frame-level synchronization during acquisition, thereby achieving collaborative input of multi-source data and providing a data foundation for subsequent defogging and depth estimation processing.

[0022] S2. Add a Scale-Cooperative Attention (SCA) module to the u-net network architecture, and introduce an Error-AwareGuidance Module (EAGM) to the downstream task.

[0023] The u-net network architecture includes three encoders and two decoders. Each encoder and decoder includes a ResBlock block. A scale-coordinated attention module is added to the residual of each ResBlock block, and only the first attention module is used to receive depth information from the feature map.

[0024] The scale-coordinated attention module includes a cross-scale tensor encoding (CSTE) module and a multi-scale encoding module.

[0025] S3. In the upstream task, a cross-scale tensor coding module is first introduced to independently slice and store features from different branches, effectively solving the scale confusion problem in traditional image dehazing scale fusion methods. The computation steps for the cross-scale tensor coding module are as follows: The input is represented as ,in This represents the input to the cross-scale tensor coding module; R Indicates the number of scales; B This indicates the batch size, which is the number of samples input into the network at one time, used for parallel computing; C H represents the number of channels; H represents the height of the feature map; W represents the width of the feature map.

[0026] First, features at different scales are extracted through three parallel branches: Local Feature Branch: In the formula: This represents the extracted local features; Represents the linear rectification activation function; express Convolution operation; This represents a linear transformation used for feature mapping or dimensionality reduction in the channel dimension; This represents the structure of a four-dimensional feature tensor.

[0027] Contextual Feature Branch: In the formula: Representing contextual features, its tensor size is ; This represents a tensor that upsamples to the original image size, representing contextual features. The tensor size is... ; This indicates a size that is upsampled to twice the original size.

[0028] Global Feature Branch: In the formula: Represents global features, and its tensor size is ; This represents the tensor representing global features after upsampling to the original image size, and its size is... ; This indicates a size that is upsampled four times to the original size.

[0029] Next, a new scale dimension D is introduced to store the features of each branch in independent slices. Specifically, the extracted local, contextual, and global features are first assigned to their respective independent slices along the D dimension, thus ensuring that features at different scales at the same spatial location remain independent of each other, as shown below: Subsequently, these individual slices are pieced together along the D dimension to form a cross-scale tensor: In the formula: Represents a multi-scale tensor; , , These represent local features, contextual features, and global features, respectively. This indicates that three features at different scales are stored independently on the new scale dimension D; This indicates the size of the newly added dimension, used to store slices of three different scales.

[0030] Here, each scale feature is preserved in an independent slice. D=1 corresponds to a local feature slice; D=2 corresponds to a contextual feature slice; and D=3 corresponds to a global feature slice. Through this step-by-step construction method, even when stitching along the new scale dimension D, the features of each branch can still be stored independently without mixing, thus explicitly preserving scale differences and mitigating scale confusion. The modeling idea of ​​this module is as follows: Figure 2 As shown.

[0031] Secondly, the introduction of a Multi-Scale Weaving Module (MSWM) enables the network to learn the interaction and complementarity of features at different scales in the same spatial location, effectively solving the problem of limited scale coordination ability in traditional multi-scale feature learning. While the cross-scale tensor encoding module ensures that multi-scale features (local, contextual, and global) are stored in independent slices, simple stacking does not fully utilize the complementary information between them. Therefore, a multi-scale 3D dilated convolution is designed to enhance inter-scale collaboration. Specifically, the kernel sizes are set to 3×3×3, 3×5×5, and 3×7×7, the dilation rate is set to 2, and these convolutions are applied in parallel on the cross-scale tensor. The output of the cross-scale tensor encoding module is represented as follows: The following is an example of applying three parallel 3D dilation convolutions with different receptive field tensors: ) In the formula: , , These represent the three branches of parallel multi-scale 3D convolution (e.g., Figure 5 (as shown) The symbols represent 3D convolution, with subscripts 3×3×3, 3×5×5, and 3×7×7 indicating the size of the 3D convolution kernel. (in, and This refers to the height and width of the 3D convolution kernel. (Represents the size of the 3D convolutional augmentation dimension). This represents the dilation rate of a 3D convolution.

[0032] Unlike standard 2D convolutions, which operate only on a single feature plane, 3D convolutions extract features along an introduced new scale dimension D, enabling the learning of differences, complementarities, and stepwise relationships between features at multiple scales within the same spatial location. The final output of the multi-scale compilation module is represented as follows: In the formula: This represents the final output of the three branches of the parallel multi-scale 3D convolution; This means reducing the dimensionality of the image in the D dimension to the original image size using average pooling. ; This represents the final output of the scale compilation module.

[0033] Finally, utilizing multi-scale interleaved modules (module architecture such as...) Figure 5 As shown) through gating weighting and coordinate attention (module architecture as follows) Figure 4 Combined with the above, a scale-coordinated attention module is formed (the module architecture is as shown). Figure 3 (as shown) Coordinate attention provides accurate spatial and channel context modeling, enabling the network to highlight important regions and capture long-range dependencies, thereby further enhancing the collaborative learning and progressive interaction of multi-scale features.

[0034] S4. Introduce an Error-Aware Guidance Mechanism (EAGM) in downstream tasks to enhance multimodal collaborative defogging capabilities through global guidance and local error awareness.

[0035] The module architecture of the error-aware guidance module is as follows: Figure 6 As shown, the Error-AwareCross Attention (EACA) module emphasizes regions with inconsistent or difficult-to-process structures from a global perspective, and integrates the local differences captured by the Local Difference Module (LDM) through a gated fusion mechanism, thereby achieving accurate feature optimization.

[0036] The following details the global feature guidance module and the local difference module, and explains how they are integrated through a gating fusion mechanism to form a complete error-aware guidance module.

[0037] Global Feature Guidance Module: The global feature guidance module takes the features of the dehazed image (or dehazed depth map) and the corresponding real haze-free image (or its depth map) as input, as shown below: In the formula: This represents the dehazed image output by the model; This represents a true, fog-free image.

[0038] The dehazed features are projected onto the query (Query, Q) through a 1×1 convolution, while the true features are projected onto the key (Key, K) and value (Value, V) through independent 1×1 convolutions, as shown below: In the formula: This represents the calculation result of cross-attention; , and These represent the query, key, and value generated from the dehazed graph and the true haze-free graph, respectively; to further expand the receptive field, multi-head attention is introduced. This represents multiple attention heads that divide the feature map; This indicates an operation that splices together the attention of multiple parties; Represents a nonlinear activation function; and These represent two different linear layers; Presentation layer normalization operation; This represents the normalized output (for numerical constraint stability). Indicates average pooling; express transpose; This represents the normalization factor, used to prevent the attention distribution from becoming too sharp; This represents the normalization function.

[0039] Then, a 3×3 convolution is applied to it, followed by a softmax normalization function to generate the final weight map. , means as follows: In the formula: This represents the generated weighted graph; Indicates a feedforward fully connected layer; This represents the normalization function.

[0040] Regions with high similarity will receive higher weights, while regions with inconsistent structures or that are difficult to process will be assigned lower weights.

[0041] The weight map obtained at this time Representing the similarity matrix, this weight map encodes the structural consistency and global feature relationships between dehazed and ground truth features. Since cross-attention is applied to dehazed and ground truth images, regions with good dehazing (high similarity) receive higher weights, while regions with poor dehazing (structural inconsistencies or difficult-to-process) are assigned lower weights. Because the depth estimation loss function needs to be weighted using poorly dehazed regions, the weight map... By reversing the process, areas with poor dehazing performance are highlighted, and the weight of these areas is amplified, as shown below: In the formula: This represents the error perception weight of the dehazed image (i.e., the weight of the region where the dehazing is not ideal).

[0042] Similarly, by inputting the dehazed depth map and the corresponding true fog-free depth map, an error-aware weight for depth estimation can be obtained. This is used to highlight areas where depth estimation in the dehazed image is inadequate. Finally, under the weighting of the loss function, this weight guides a collaborative dehazing process based on depth information.

[0043] Let the dehazed image be The real, fog-free image is The image containing fog is Their corresponding depth maps are respectively and The error-aware weights generated by the global feature-guided module for the dehazed image and depth estimation are denoted as follows: = , The loss functions of the depth estimation network and the defogging network are denoted as follows: and Based on these definitions, the collaborative dehazing guided by the global feature guidance module can be expressed as: In the formula: This represents the loss function of the depth estimation network guided by the global feature-driven module; This represents the loss function of the dehazing network guided by global features. By weighting each other's loss functions on their respective undesirable regions, the two networks can focus on the regions they are not optimizing well, thus achieving collaborative dehazing.

[0044] exist and Guided by both structural and global perspectives, the depth estimation task focuses not only on regions with inconsistent structures in the dehazed image but also on regions with mismatched global feature relationships. Meanwhile, the dehazing task emphasizes regions where depth estimation is unreliable or lacks global consistency. This design achieves complementary optimization at both the structural and global levels, resulting in more robust image restoration.

[0045] The Local Difference module aims to extract local inconsistencies between the dehazed image and the corresponding real haze-free image (or depth map), thereby highlighting areas where dehazing is more difficult. It serves as a fine-grained supplement to the global guidance provided by the global feature guidance module.

[0046] set up This indicates the weight of areas with poor defogging performance. These represent the weights for depth estimation in these regions. The specific calculation steps of the local difference module are as follows: In the formula: Represents a multilayer perceptron; Represents the normalization function; This represents a 3×3 two-dimensional convolution.

[0047] Subsequently, the collaborative dehazing guided by the Local Difference Module (LDM) can be represented as: In the formula: and Let represent the loss functions of the dehazing network and the depth estimation network guided by the local difference module, respectively.

[0048] In this stage, collaborative dehazing is guided globally by a global feature guidance module and modeled by a local difference module (LDM) that detects local inconsistencies. Finally, the weights of the two modules are gated and weighted to form the final error-aware guidance module.

[0049] Let the weight of the area with poor final defogging effect be . , The weights for depth estimation in these regions, and the final performance of dehazing using depth information assisted by the Error Aware Guidance Module (EAGM), are represented as follows: In the formula: and These represent the loss functions of the dehazing network and the depth estimation network guided by the error-aware guidance module, respectively.

[0050] S5. The system performs post-processing and optimization on the results after defogging and depth estimation, and displays the clear image and corresponding depth information on the display terminal. At the same time, the result data is stored in the database or local storage module for subsequent safety monitoring, personnel identification and auxiliary rescue.

[0051] The system outputs the dehazed images and depth estimation results processed by the scale-coordinated attention and error-aware guidance module. This module can display the restored haze-free images and corresponding depth maps in real time on the monitoring terminal, allowing operators to intuitively observe the field environment. Simultaneously, the system supports automatically storing the results data and related processing logs to a local database or remote server, facilitating subsequent backtracking, analysis, and model iteration optimization. The output interface can be expanded to multi-terminal display and mobile device access as needed, ensuring rapid acquisition and scheduling of image information in complex scenarios such as deep-earth medical rescue, thereby improving the system's practicality and reliability.

[0052] The results of this depth-guided image dehazing method for deep-earth medical rescue compared with existing algorithms such as OK-Net (Omni-Kernel Network for Image Restoration), MIT-Net (Mutual information-driven triple interaction network for efficient image dehazing), DehazeFormer (Image dehazing transformer with transmission-aware 3D position embedding), and DCMPNet (Depth Information Assisted Collaborative Mutual Promotion Network for Single Image Dehazing) are shown in the figure below. Figure 7 As shown (where Hazy and GT represent the haze image and the ground truth image, respectively). (By...) Figure 7It is evident that the algorithm proposed in this invention has stronger performance and potential in the field of image dehazing. This is due to its scale-coordinated attention, which can effectively solve the problems of scale aliasing and weak cross-scale synchronous extraction capabilities in multi-scale feature learning. Combined with the depth estimation network, it can effectively help the network find the fog concentration distribution area. Furthermore, by combining the error perception mechanism, through global guidance and local difference guidance, the dehazing and depth estimation networks can focus on the undesirable areas and perform further collaborative optimization for dehazing.

[0053] This depth-guided image dehazing method for deep-earth medical rescue aims to improve the visibility and accuracy and real-time performance of structural information restoration in medical rescue monitoring images in deep-earth environments such as mines. The method focuses on monocular image dehazing, combined with depth estimation. It introduces a scale-coordinated attention module to achieve collaborative learning of multi-scale features, mitigating scale confusion and enhancing cross-scale feature collaboration. Simultaneously, an error-aware guidance module fuses local feature differences and global dependencies, mutually optimizing dehazing and depth estimation tasks. This results in more accurate image restoration in areas with complex structures or uneven fog distribution. This depth-guided image dehazing method for deep-earth medical rescue achieves close collaborative optimization and performance improvement between image dehazing and depth estimation tasks. It can, to some extent, address the problems of scale confusion and limited scale coordination in multi-scale feature learning, thereby improving the accuracy of monitoring camera systems and intelligent detection systems in mines and enhancing coal mine safety.

Claims

1. A method for dehazing depth-guided images for deep-earth medical rescue, characterized in that, Specifically, the following steps are included: S1. The system acquires foggy images in mines or deep underground environments using an industrial camera, and simultaneously uses a depth sensor to obtain optional depth information, providing input data for subsequent defogging and depth estimation. S2. Add a scale-coordinated attention module to the u-net network architecture, and introduce an error-aware guidance module in the downstream tasks. S3. In the upstream task, a cross-scale tensor coding module is first introduced to independently slice and store the features of different branches. Then, a multi-scale programming module is introduced to enable the network to learn the interaction and complementarity of features at different scales in the same spatial location; finally, a scale co-attention module is formed by combining the multi-scale interleaving module with gating weighting and coordinate attention. S4. Introduce an error-aware guidance module in downstream tasks to enhance multimodal collaborative defogging capabilities through global guidance and local error awareness; S5. The system performs post-processing and optimization on the results after defogging and depth estimation, and displays the clear image and corresponding depth information on the display terminal. At the same time, the result data is stored in the database or local storage module for subsequent safety monitoring, personnel identification and auxiliary rescue.

2. The method for dehazing depth-guided images for deep-earth medical rescue according to claim 1, characterized in that, The u-net network architecture in S2 includes three encoders and two decoders. Each encoder and decoder includes a ResBlock block. A scale-coordinated attention module is added to the residual of each ResBlock block, and only the first attention module is used to receive the depth information of the feature map.

3. The method for dehazing depth-guided images for deep-earth medical rescue according to claim 2, characterized in that, The scale-coordinated attention module in S2 includes a cross-scale tensor coding module and a multi-scale coding module.

4. The method for dehazing depth-guided images for deep-earth medical rescue according to claim 3, characterized in that, The calculation steps for the S3 cross-scale tensor coding module are as follows: The input is represented as ,in This represents the input to the cross-scale tensor coding module. Indicates the scale quantity. Indicates batch size, Indicates the number of channels. Indicates the height of the feature map. Indicates the width of the feature map; First, features at different scales are extracted through three parallel branches: Local feature branches: In the formula: This represents the extracted local features; Represents the linear rectification activation function; express Convolution operation; Represents a linear transformation; Regional feature branches: In the formula: Represents contextual features; This represents a tensor that upsamples the image to the original image size, representing the contextual features. This indicates a size that is upsampled to twice the original size; Global feature branch: In the formula: Represents global features; This represents a tensor that upsamples the image to the original image size, representing global features. This indicates a size upsampled four times to the original size; Secondly, a new scale dimension D is introduced, and the extracted local, contextual, and global features are first assigned to their respective independent slices along the D dimension, as shown below: Finally, the individual slices are stitched together along the D dimension to form a cross-scale tensor, as shown below: In the formula: Represents a multi-scale tensor; This indicates that three features at different scales are stored independently on the new scale dimension D; Indicates the size of the newly added dimension; Each scale feature is preserved in an independent slice, with D=1 corresponding to the local feature slice, D=2 corresponding to the context feature slice, and D=3 corresponding to the global feature slice.

5. The method for dehazing depth-guided images for deep-earth medical rescue according to claim 4, characterized in that, The calculation steps of the S3 multi-scale compilation module are as follows: Design a multi-scale 3D dilated convolution, setting the kernel size to 3×3×3, 3×5×5, and 3×7×7, and the dilation rate to 2. Apply these convolutions in parallel across the multi-scale tensors. The output of the multi-scale tensor encoding module is represented as follows: The output of the cross-scale tensor coding module is applied with three parallel 3D dilated convolutions of tensors with different receptive fields, as shown below: In the formula: , , These represent the three branches of parallel multi-scale 3D convolution; The symbols represent 3D convolution, with subscripts 3×3×3, 3×5×5, and 3×7×7 indicating the size of the 3D convolution kernel. ; This represents the dilation rate of a 3D convolution; The final output of the multi-scale compilation module is represented as follows: In the formula: This represents the final output of the three branches of the parallel multi-scale 3D convolution; This means reducing the dimensionality of the image in the D dimension to the original image size using average pooling. This represents the final output of the scale compilation module.

6. The method for dehazing depth-guided images for deep-earth medical rescue according to claim 5, characterized in that, The error-aware guidance module of S4 emphasizes regions with inconsistent or difficult-to-handle structures from a global perspective through the global feature guidance module, and integrates the local differences captured by the local difference module through the gating fusion mechanism to achieve feature optimization.

7. The method for dehazing depth-guided images for deep-earth medical rescue according to claim 6, characterized in that, The specific working process of the error-aware guidance module of S4 is as follows: ① Global Feature Guidance Module: The global feature guidance module takes the features of the dehazed image or dehazed depth map and the corresponding real haze-free image or real haze-free depth map as input, as shown below: In the formula: This represents the dehazed image output by the model; Represents a true, fog-free image; The dehazed features are projected onto the query Q through a 1×1 convolution, while the true features are projected onto the key K and value V through independent 1×1 convolutions, as shown below: In the formula: This represents the calculation result of cross-attention; , and These represent the query, key, and value generated from the dehazed graph and the true haze-free graph, respectively. This represents multiple attention heads that divide the feature map; This indicates an operation that splices together the attention of multiple parties; Represents a nonlinear activation function; and These represent two different linear layers; Presentation layer normalization operation; This represents the normalized output; Indicates average pooling; express transpose; Indicates the normalization factor; Represents the normalization function; Then, a 3×3 convolution is applied, followed by a softmax normalization function to generate the final weight map. , means as follows: In the formula: This represents the generated weighted graph; Indicates a feedforward fully connected layer; For weighted graph Reversing the order is represented as follows: In the formula: Represents the error-aware weights of the dehazed image; Similarly, by inputting the dehazed depth map and the corresponding true fog-free depth map, the error-aware weights for depth estimation are obtained. ; Let the dehazed image be The real, fog-free image is The image containing fog is , , and The corresponding depth maps are as follows: and The error-aware weights generated by the global feature guidance module for the dehazed image and depth estimation are denoted as follows: = , The collaborative dehazing guided by the global feature guidance module is then represented as: In the formula: This represents the loss function of the depth estimation network guided by the global feature-driven module; This represents the loss function of the dehazing network guided by the global feature module; ②Local difference module: set up This indicates the weight of areas with poor defogging performance. The weights for depth estimation in areas with poor dehazing performance are represented by the following steps in the local difference module: In the formula: Represents a multilayer perceptron; Represents the normalization function; This represents a 3×3 two-dimensional convolution; The collaborative dehazing guided by the local difference module is then represented as: In the formula: and These represent the loss functions of the dehazing network and the depth estimation network guided by the local difference module, respectively. ③ Collaborative dehazing provides global guidance through a global feature guidance module and uses a local difference module to detect local inconsistencies for modeling. The weights of the two modules are gated and weighted to form the final error-aware guidance module: Let the weight of the area with poor final defogging effect be . , The weights for depth estimation in areas with poor dehazing performance are represented as follows. The final performance of dehazing using depth information assisted by the error-aware guidance module is expressed as follows: In the formula: and These represent the loss functions of the dehazing network and the depth estimation network guided by the error-aware guidance module, respectively.

8. A depth-guided image dehazing system for deep-earth medical rescue, based on the depth-guided image dehazing method for deep-earth medical rescue as described in claim 1, characterized in that, It includes an image acquisition unit, a dehazing unit, a depth estimation unit, and a post-processing and storage module; The image acquisition unit includes an industrial camera for acquiring foggy images and a depth sensor for simultaneously acquiring corresponding depth information; The dehazing unit includes a scale-coordinated attention module, which in turn includes a cross-scale tensor coding module and a multi-scale encoding module. The depth estimation unit includes an error-aware guidance module, which in turn includes a global feature guidance module and a local difference module. The post-processing and storage module includes a monitoring terminal.

Images