Lightweight mine image super-resolution reconstruction system and method based on deep feature calibration

Abstract

The present invention belongs to the field of image resolution and reconstruction technology, and discloses a lightweight mine image super-resolution reconstruction system and method based on deep feature calibration. The reconstruction system includes a shallow feature extraction layer, a deep feature extraction layer, and a reconstruction module layer. The reconstruction method is as follows: shallow features are extracted through the shallow feature extraction layer, and the shallow features are input into multiple stacked residual attention feature modules for deep feature extraction. The deep features are first aggregated, and then fused with the shallow features in step 1. The overlapped image is input into the reconstruction module layer for image reconstruction to obtain a super-resolution image. The present invention obtains more accurate results in image reconstruction and has better reconstruction effects for mine images with complex structures.

Description

Technical Field

[0001] The present invention belongs to the technical field of image resolution and reconstruction, and in particular relates to a lightweight mine image super-resolution reconstruction system and method based on depth feature calibration. Background Art

[0002] Image super-resolution technology plays a vital role in fields such as medical imaging, smart mining, and computational photography. It aims to upscale low-resolution images to high resolution, enhancing image clarity and detail. Currently, the main super-resolution reconstruction methods include interpolation, reconstruction, and learning.

[0003] Interpolation methods generally employ bicubic interpolation and geometric feature-based interpolation. While computationally simple and fast, the resulting images suffer from significant deficiencies in detail and realism, and are prone to artifacts such as image oscillation and edge blurring. Reconstruction methods, such as maximum a posteriori estimation and iterative backprojection, leverage prior knowledge and statistical properties of images to attempt to infer the likely form of a high-resolution image from a low-resolution image. However, these methods are computationally complex and sensitive to the upscaling factor; different upscaling factors may require readjustment of model parameters, which presents certain limitations in practical applications. With the increasing maturity of machine learning techniques, learning methods are beginning to gain prominence in the field of image super-resolution. Neighborhood embedding methods map low-resolution image patches to neighborhoods in the high-resolution image space, leveraging the similarity between pairs of samples to reconstruct high-resolution patches. While this method can capture local image features to a certain extent, it has limited capabilities for complex scenes and large-scale images. Sparse representation methods, based on the assumption of image sparsity under a specific dictionary, achieve high-resolution image reconstruction by learning sparse representation coefficients for low-resolution and high-resolution image pairs. It shows certain advantages in processing image textures and details, but also faces problems of computational efficiency and model adaptability.

[0004] Chinese patent application ZL CN2023111940240 discloses a lightweight image super-resolution reconstruction system and method for edge mobile devices, in which the model can better utilize multi-scale information and improve the perception of details and overall structure. However, it does not continuously focus on more detailed information such as edges and textures of the original low-resolution image, and there is a tendency to ignore or lose high-frequency information.

[0005] Chinese patent application ZL CN2023110111616 discloses a super-resolution reconstruction method for mine images based on residual mixed attention. The proposed multi-scale residual dense mixed attention super-resolution network pursues the improvement of indicators, but still has problems such as not paying attention to whether the reconstructed image is consistent with human perception, whether it is close to the true image, and insufficient accuracy of the reconstructed image. At the same time, the network model in this patent application directly stacks the number of network modules, which improves the performance of the network while causing huge computational overhead, and is not suitable for application in mobile terminals or edge computing devices in mine environments.

[0006] In summary, current super-resolution networks often focus on image reconstruction performance while neglecting their real-world applications on mobile devices. Mine video surveillance, in particular, plays a crucial role in safety monitoring and alarms. However, due to the unique environment and equipment limitations of mines, images captured often suffer from quality and visual issues. Image super-resolution technology, which can reconstruct mine images and restore them to clear, high-quality images, is of great significance. The computational cost of super-resolution models is closely related to the number of modules and their architecture. As network depth increases, the network fails to fully utilize information from all convolutional layers and feature information at varying granularities, resulting in insufficient richness in the reconstructed image. Furthermore, lightweight layers often lack global information guidance, resulting in a discrepancy between the reconstructed image and the real image, which is inconsistent with human visual perception. Summary of the Invention

[0007] In order to solve the above technical problems, the present invention provides a lightweight mine image super-resolution reconstruction system and method based on deep feature calibration, designs a deep feature calibration block and deep feature calibration attention, and on this basis, constructs a lightweight mine image super-resolution reconstruction system based on deep feature calibration. The system has better reconstruction effect for mine images with complex structures.

[0008] In order to achieve the above object, the present invention is achieved through the following technical solutions:

[0009] The present invention is a lightweight mine image super-resolution reconstruction system based on deep feature calibration. The lightweight mine image super-resolution reconstruction system includes a shallow feature extraction layer, a deep feature extraction layer and a reconstruction module layer, wherein:

[0010] The shallow feature extraction layer is a convolutional layer with a kernel size of 3×3, which performs shallow feature extraction on the input image, and the output of the shallow feature extraction layer serves as the input of the deep feature extraction layer;

[0011] The deep feature extraction layer is composed of multiple residual attention feature modules (FARGs), which include a deep feature calibration attention module (HFFCA) and a deep feature calibration module. The deep feature calibration module includes a high-frequency feature refinement module (HFFOB), a high-frequency feature non-redundant module (HFNRB), and a multi-scale feature enhancement module (MSFEB).

[0012] The reconstruction module layer includes a sub-pixel convolution layer and a convolution layer with a kernel size of 3×3, and the sub-pixel convolution layer includes a convolution layer with a kernel size of 3×3.

[0013] The present invention provides a lightweight mine image super-resolution reconstruction method based on depth feature calibration, and the reconstruction method specifically includes the following steps:

[0014] Step 1: Perform shallow feature extraction on the input image through the shallow feature extraction layer to obtain extracted shallow features;

[0015] Step 2: Input the shallow features extracted in step 1 into multiple stacked residual attention feature modules (FARG) for deep feature extraction;

[0016] Step 3: Aggregate the output of each residual attention feature module (FARG) in step 2 and fuse it with the shallow features extracted in step 1;

[0017] Step 4: Input the features fused in step 3 into the image reconstruction module layer for image reconstruction, and output the reconstructed super-resolution image.

[0018] A further improvement of the present invention is that: in step 1, shallow feature extraction is performed on the input image through the shallow feature extraction layer, and the expression is: F0=Conv3(I LR );

[0019] Among them, I LR is the input image; F0 is the extracted shallow feature; Conv3(·) is the convolution operation with a kernel size of 3×3.

[0020] A further improvement of the present invention is that in step 2, the shallow features extracted are input into multiple stacked residual attention feature modules (FARG) for deep feature extraction, which specifically includes the following steps:

[0021] Step 2.1: Input the extracted shallow features into multiple cascaded deep feature calibration modules to fully extract high-frequency information that is conducive to image texture detail recovery, such as texture details and edge features, while ensuring the non-redundancy of high-frequency information;

[0022] Step 2.2: The output of each deep feature calibration module is input into the deep feature calibration attention module (HFFCA) along the transmission path of the residual attention feature module to calibrate and select high-frequency information that is critical for image texture detail recovery;

[0023] Step 2.3, fuse the output of each deep feature calibration module with the output of the deep feature calibration attention module (HFFCA), and then pass it through channel contrast attention (CCA) to capture image features of different channel dimensions and obtain richer feature information. The fusion is divided into feature aggregation path and attention aggregation path.

[0024] A further improvement of the present invention is that in step 2.1, the deep feature calibration module extracts high-frequency information that is beneficial to image texture detail recovery through a high-frequency feature refinement module (HFFOB), a high-frequency feature non-redundant module (HFNRB), and a multi-scale feature enhancement module (MSFEB), while ensuring the non-redundancy of the high-frequency information, specifically comprising the following steps:

[0025] Step 2.1.1, the high-frequency feature refinement module (HFFOB) uses different multiples of upsampling and downsampling methods, standard convolution and depth-wise separable convolution to extract high-frequency information, which is expressed as:

[0026] Y1=H LFEB (H Sampling (X1))

[0027] Y2=SConv(Gelu(DeConv(X1)))

[0028]

[0029] Among them, H Sampling (·) is the function of upsampling and downsampling fusion after double, quadruple, and eightfold sampling, H LFEB (·) is the function of LFEB, DeConv(·) is the 2×2 deconvolution operation, SConv(·) is the strided convolution with a step size of 2, σ is the Sigmoid function, It is expressed as element-by-element multiplication, where Y is the output of the high-frequency feature refinement module HFFOB, and Y1 and Y2 are temporary tensors in the neural network;

[0030] Step 2.1.2, the high-frequency feature non-redundant module (HFNRB) uses maximum pooling, standard convolution, and depth-wise separable convolution to reduce repeated high-frequency information in feature propagation to ensure the non-redundancy of high-frequency information, which is expressed as:

[0031] Z1=H LEFB (X2′)

[0032] Z2=Gelu(Conv1(H Max (X″2)))

[0033] Z=Conv1(Cat(Z1+Z2))+X2

[0034] Among them, H Max (·) is the maximum pooling function, Z is the output of the high-frequency feature non-redundant module HNFRB, Z1 and Z2 are temporary tensors in the neural network;

[0035] Step 2.1.3: The multi-scale feature enhancement module (MSFEB) guides and learns the global information of the features aggregated by the high-frequency feature refinement module (HFFOB) and the high-frequency feature non-redundant module (HFNRB) to obtain comprehensive and rich output features of the multi-scale feature enhancement module (MSFEB), which can be expressed as:

[0036] M1=Conv1(H CS (Cat(Y,Z)))

[0037]

[0038] M=DW 3×3 (Conv1(M2))+Gelu(DW 3×3 (Conv1(M2)))+X2

[0039] Among them, H CS (·) is the channel shuffle operation, DW m×n (·) is a depthwise separable convolution with kernel size m×n, M is the output of the multi-scale feature enhancement module (MSFEB), and M1 and M2 are temporary tensors in the neural network.

[0040] A further improvement of the present invention is that the step 2.2 specifically includes the following steps:

[0041] Step 2.2.1. Use the output of the deep feature calibration module as the input feature, perform a 1×1 convolution operation on the input feature, and then perform a channel splitting operation to divide the input feature into two parts;

[0042] Step 2.2.2: Use the parallel branches of max pooling and average pooling to input the features of step 2.2.1 into the max pooling branch and average pooling branch respectively, effectively highlighting the information area and helping the model learn the main features of the image;

[0043] Step 2.2.3: Use a convolutional layer with a kernel size of 1×1 and a Sigmoid activation function to multiply the outputs of the two branches with the corresponding inputs to perform feature high-frequency calibration.

[0044] Step 2.2.4: The outputs of the two branches are fused together through a convolutional layer with a kernel size of 1×1 and Concat splicing to form the output of the attention feature.

[0045] A further improvement of the present invention is that: in step 2.3, the output of each deep feature calibration module and the output of the deep feature calibration attention module (HFFCA) are fused, specifically: in the feature aggregation path, in order to effectively utilize the feature information, the feature output of each deep feature calibration module (HFFCB) is sent to the end of RAFG, and feature aggregation is performed. In the attention aggregation path, the attention output of each deep feature calibration module is directly sent to the deep feature calibration attention module to calibrate and select high-frequency information that is more critical to the image texture details, and then transmitted to the end of RAFG for aggregation. Finally, the channel contrast attention CCA is used to adjust the feature information after the fusion of the feature aggregation path and the attention aggregation path, so that it focuses on the key and important channel content, and then added to the initial input.

[0046] A further improvement of the present invention is that step 3 aggregates the output of each residual attention feature module and fuses it with the shallow features; it is expressed as:

[0047]

[0048] Among them, F FC n ,F AC n represents the feature output and attention output of the nth residual attention feature module, Concat(·) represents the cascade on the channel dimension; Conv1(·) is a 1×1 convolution, F ff Represents the output after feature fusion.

[0049] A further improvement of the present invention is that: the step 4 inputs the fused features into the image reconstruction module layer for image reconstruction, and outputs the reconstructed super-resolution image I SR , expressed as:

[0050] I SR =Conv3(H PixelShuffle (F ff +F0))

[0051] Among them, H PixelShuffle (·) is the sub-pixel convolution layer, I SR is the reconstructed super-resolution image.

[0052] The beneficial effects of the present invention are:

[0053] The present invention designs a deep feature calibration block, which fully extracts high-frequency information that is beneficial to the recovery of image texture details through high-frequency feature refinement blocks and high-frequency feature non-redundant blocks, while ensuring the non-redundancy of high-frequency information; in addition, through the multi-scale feature enhancement module, the receptive field range is further expanded, and global information is captured, thereby improving the accuracy and visual quality of image reconstruction.

[0054] This paper proposes deep feature calibration attention, which accurately calibrates and selects high-frequency information that is crucial for image texture detail recovery through high-frequency feature calibration attention, thereby obtaining more accurate results in image reconstruction.

[0055] Based on the deep feature calibration block and deep feature calibration attention, the present invention further constructs a lightweight mine image super-resolution reconstruction system based on deep feature calibration, which has better reconstruction effect for mine images with complex structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a structural framework diagram of the lightweight mine image super-resolution reconstruction system of the present invention.

[0057] Figure 2 It is a structural diagram of the depth feature calibration module of the present invention.

[0058] Figure 3 This is a visual comparison diagram of the reconstructed image using the method of the present invention and the four-fold reconstruction result of the low-resolution image on the mine image. DETAILED DESCRIPTION

[0059] The following diagrams illustrate embodiments of the present invention. For clarity, many practical details are included in the following description. However, it should be understood that these practical details are not intended to limit the present invention. In other words, in some embodiments of the present invention, these practical details are not essential.

[0060] like Figure 1As shown, the present invention is a lightweight mine image super-resolution reconstruction system based on deep feature calibration, characterized in that: the lightweight mine image super-resolution reconstruction system includes a shallow feature extraction layer, a deep feature extraction layer and a reconstruction module layer, wherein: the shallow feature extraction layer is a convolution layer with a kernel size of 3×3, which performs shallow feature extraction on the input image, and the output of the shallow feature extraction layer is used as the input of the deep feature extraction layer; the deep feature extraction layer is composed of multiple residual attention feature modules (FARG), and the residual attention feature module (FARG) includes a deep feature calibration attention module (HFFCA) and a deep feature calibration module, and the deep feature calibration module includes a high-frequency feature refinement module (HFFOB), a high-frequency feature non-redundant module (HFNRB), and a multi-scale feature enhancement module (MSFEB); the reconstruction module layer includes a sub-pixel convolution layer and a convolution layer with a kernel size of 3×3, and the sub-pixel convolution layer includes a convolution layer with a kernel size of 3×3.

[0061] When used, the low-resolution image is input into the shallow feature extraction module, and the low-resolution image I LR After a 3×3 convolution, shallow features are extracted to obtain shallow features F0; shallow features F0 are processed by multiple residual attention feature modules to obtain deep features F n ; Deep feature F n After adding it to the shallow feature F0 and passing through the sub-pixel convolution layer and 3×3 convolution, the reconstructed super-resolution image I is obtained. SR .

[0062] Based on the reconstruction system, the present invention also proposes a lightweight mine image super-resolution reconstruction method based on depth feature calibration, and the reconstruction method specifically includes the following steps:

[0063] Step 1: Perform shallow feature extraction on the input image through the shallow feature extraction layer to obtain the extracted shallow features, which are expressed as:

[0064] F0=Conv3(I LR );

[0065] Among them, I LR is the input image; F0 is the extracted shallow feature; Conv3(·) is the convolution operation with a kernel size of 3×3.

[0066] Step 2: Input the shallow features extracted in step 1 into multiple stacked residual attention feature modules (FARG) for deep feature extraction. The specific steps include the following:

[0067] Step 2.1: Input the extracted shallow features into multiple cascaded deep feature calibration modules to fully extract high-frequency information that is conducive to image texture detail recovery, such as texture details and edge features, while ensuring the non-redundancy of high-frequency information. The specific steps include the following:

[0068] Step 2.1.1, the high-frequency feature refinement module (HFFOB) uses different multiples of upsampling and downsampling methods, standard convolution and depth-wise separable convolution to extract high-frequency information, which is expressed as:

[0069] Y1=H LFEB (H Sampling (X1))

[0070] Y2=SConv(Gelu(DeConv(X1)))

[0071]

[0072] Among them, H Sampling (·) is the function of upsampling and downsampling fusion after double, quadruple, and eightfold sampling, H LFEB (·) is the function of LFEB, DeConv(·) is the 2×2 deconvolution operation, SConv(·) is the strided convolution with a step size of 2, σ is the Sigmoid function, It is represented as element-by-element multiplication, where Y is the output of HFFOB, and Y1 and Y2 are temporary tensors in the neural network;

[0073] Step 2.1.2, the high-frequency feature non-redundant module (HFNRB) uses maximum pooling, standard convolution, and depth-wise separable convolution to reduce repeated high-frequency information in feature propagation to ensure the non-redundancy of high-frequency information, which is expressed as:

[0074] Z1=H LEFB (X′2)

[0075] Z2=Gelu(Conv1(H Max (X″2)))

[0076] Z=Conv1(Cat(Z1+Z2))+X2

[0077] Among them, H Max (·) is the maximum pooling function, Z is the output of HNFRB, Z1 and Z2 are temporary tensors in the neural network;

[0078] Step 2.1.3: The multi-scale feature enhancement module (MSFEB) guides and learns the global information of the features aggregated by the high-frequency feature refinement module (HFFOB) and the high-frequency feature non-redundant module (HFNRB) to obtain comprehensive and rich output features of the multi-scale feature enhancement module (MSFEB), which can be expressed as:

[0079] M1=Conv1(H CS (Cat(Y,Z)))

[0080]

[0081] M=DW 3×3 (Conv1(M2))+Gelu(DW 3×3 (Conv1(M2)))+X2

[0082] Among them, H CS (·) is the channel shuffle operation, DW m×n (·) is a depthwise separable convolution with kernel size m×n, M is the output of the multi-scale feature enhancement module (MSFEB), and M1 and M2 are temporary tensors in the neural network.

[0083] Step 2.2: Input the output of each deep feature calibration module into the deep feature calibration attention module (HFFCA) along the transmission path of the residual attention feature module to calibrate and select high-frequency information that is critical for image texture detail recovery. The specific steps include the following:

[0084] Step 2.2.1. Use the output of the deep feature calibration module as the input feature, perform a 1×1 convolution operation on the input feature, and then perform a channel splitting operation to divide the input feature into two parts;

[0085] Step 2.2.2: Use the parallel branches of max pooling and average pooling to input the features of step 2.2.1 into the max pooling branch and average pooling branch respectively, effectively highlighting the information area and helping the model learn the main features of the image;

[0086] Step 2.2.3: Use a convolutional layer with a kernel size of 1×1 and a Sigmoid activation function to multiply the outputs of the two branches with the corresponding inputs to perform feature high-frequency calibration.

[0087] Step 2.2.4: The outputs of the two branches are fused together through a convolutional layer with a kernel size of 1×1 and Concat splicing to form the output of the attention feature.

[0088] Step 2.3, the output of each deep feature calibration module is fused with the output of the deep feature calibration attention module (HFFCA), and then it is passed through the channel contrast attention (CCA) to capture the image features of different channel dimensions and obtain richer feature information. The fusion is divided into a feature aggregation path and an attention aggregation path, and the output of each deep feature calibration module is fused with the output of the deep feature calibration attention module (HFFCA). Specifically, in the feature aggregation path, in order to effectively utilize the feature information, the feature output of each deep feature calibration module (HFFCB) is sent to the end of RAFG, and feature aggregation is performed. In the attention aggregation path, the attention output of each deep feature calibration module is directly sent to the deep feature calibration attention module to calibrate and select high-frequency information that is more critical to the image texture details, and then transmitted to the end of RAFG for aggregation. Finally, the channel contrast attention CCA is used to adjust the feature information after the fusion of the feature aggregation path and the attention aggregation path, so that it focuses on the key and important channel content, and then added to the initial input.

[0089] Step 3: Aggregate the output of each residual attention feature module (FARG) in step 2 and fuse it with the shallow features extracted in step 1, expressed as:

[0090]

[0091] Among them, F FC n ,F AC n represents the feature output and attention output of the nth residual attention feature module, Concat(·) represents the cascade on the channel dimension; Conv1(·) is a 1×1 convolution, F ff Represents the output after feature fusion.

[0092] Step 4: Input the features fused in step 3 into the image reconstruction module layer for image reconstruction, and output the reconstructed super-resolution image I SR , expressed as:

[0093] I SR =Conv3(H PixelShuffle (F ff +F0))

[0094] Among them, H PixelShuffle (·) is the sub-pixel convolution layer, I SR is the reconstructed super-resolution image.

[0095] In order to evaluate the feasibility and effectiveness of the method of the present invention, this embodiment was experimentally verified, and the specific contents are as follows:

[0096] like Figure 3 To further verify the superiority of the proposed super-resolution image reconstruction method, a 4x image reconstruction experiment was conducted on a dataset of mine images with complex textures. The images reconstructed using this method were visually compared with those from low-resolution images. These image results clearly show that the proposed method can reconstruct images more realistically than low-resolution images, and the reconstructed images are essentially consistent with the original images. This demonstrates the significant superiority of the proposed method in image super-resolution reconstruction.

[0097] The foregoing is merely an embodiment of the present invention and is not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. A lightweight mine image super-resolution reconstruction system based on deep feature calibration, characterized by: The lightweight mine image super-resolution reconstruction system includes a shallow feature extraction layer, a deep feature extraction layer and a reconstruction module layer, wherein: The shallow feature extraction layer is a convolutional layer with a kernel size of 3×3, which performs shallow feature extraction on the input image, and the output of the shallow feature extraction layer serves as the input of the deep feature extraction layer; The deep feature extraction layer is composed of multiple residual attention feature modules (FARGs), which include a deep feature calibration attention module (HFFCA) and a deep feature calibration module (HFFCB). The deep feature calibration module includes a high-frequency feature refinement module (HFFOB), a high-frequency feature non-redundant module (HFNRB), and a multi-scale feature enhancement module (MSFEB). The reconstruction module layer includes a sub-pixel convolution layer and a convolution layer with a kernel size of 3×3, and the sub-pixel convolution layer includes a convolution layer with a kernel size of 3×3, wherein: The extracted shallow features are input into multiple stacked residual attention feature modules (FARG) for deep feature extraction, which specifically includes the following steps: Step 2.1: Input the extracted shallow features into multiple cascaded deep feature calibration modules to extract high-frequency information that is beneficial to image texture detail recovery while ensuring the non-redundancy of high-frequency information. In this step, the deep feature calibration module extracts high-frequency information that is beneficial to image texture detail recovery through a high-frequency feature refinement module (HFFOB), a high-frequency feature non-redundant module (HFNRB), and a multi-scale feature enhancement module (MSFEB), while ensuring the non-redundancy of high-frequency information. Specifically, the steps are as follows: Step 2.1.1, the high-frequency feature refinement module (HFFOB) uses different multiples of upsampling and downsampling methods, standard convolution and depth-wise separable convolution to extract high-frequency information, which is expressed as: Y1=H LFEB (H Sampling (X1)) Y2=SConv(Gelu(DeConv(X1))) Among them, H Sampling (·) is the function of upsampling and downsampling fusion after double, quadruple, and eightfold sampling, H LFEB (·) is the function of LFEB, DeConv(·) is the 2×2 deconvolution operation, SConv(·) is the strided convolution with a step size of 2, σ is the Sigmoid function, It is represented as element-by-element multiplication, where Y is the output of HFFOB, and Y1 and Y2 are temporary tensors in the neural network; Step 2.1.2, the high-frequency feature non-redundant module (HFNRB) uses maximum pooling, standard convolution, and depth-wise separable convolution to reduce repeated high-frequency information in feature propagation to ensure the non-redundancy of high-frequency information, which is expressed as: Z1=H LFEB (X2′) Z2=Gelu(Conv1(H Max (X2″))) Z=Conv1(Cat(Z1+Z2))+X2 Among them, H Max (·) is the maximum pooling function, Z is the output of HNFRB, Z1 and Z2 are temporary tensors in the neural network; Step 2.1.3: The multi-scale feature enhancement module (MSFEB) guides and learns the global information of the features aggregated by the high-frequency feature refinement module (HFFOB) and the high-frequency feature non-redundant module (HFNRB) to obtain the output features of the multi-scale feature enhancement module (MSFEB), which is expressed as: M1=Conv1(H CS (Cat(Y,Z))) <h2 style=";text-align:left;direction:ltr">M=DW<h2 style=";text-align:left;direction:ltr"> 3×3 <h2 style=";text-align:left;direction:ltr"> (Conv1(M2))+Gelu(DW<h2 style=";text-align:left;direction:ltr"> 3×3 <h2 style=";text-align:left;direction:ltr"> (Conv1(M2))+X2 Among them, H CS (·) is the channel shuffle operation, DW m×n (·) is a depth-wise separable convolution with a kernel size of m×n, M is the output of the multi-scale feature enhancement module (MSFEB), M1 and M2 are temporary tensors in the neural network; Step 2.2: Input the output of each deep feature calibration module into the deep feature calibration attention module (HFFCA) along the transmission path of the residual attention feature module to calibrate and select high-frequency information; Step 2.3, fuse the output of each deep feature calibration module and the output of the deep feature calibration attention module (HFFCA), and then pass it through channel contrast attention (CCA) to capture image features of different channel dimensions and obtain feature information, where the fusion is divided into feature aggregation path and attention aggregation path.

2. A lightweight mine image super-resolution reconstruction method based on deep feature calibration, characterized by: The lightweight mine image super-resolution reconstruction method implements the reconstruction system of claim 1, and the reconstruction method specifically includes the following steps: Step 1: Perform shallow feature extraction on the input image through the shallow feature extraction layer to obtain extracted shallow features; Step 2: Input the shallow features extracted in step 1 into multiple stacked residual attention feature modules (FARG) for deep feature extraction; Step 3: Aggregate the output of each residual attention feature module (FARG) in step 2 and fuse it with the shallow features extracted in step 1; Step 4: Input the features fused in step 3 into the image reconstruction module layer for image reconstruction, and output the reconstructed super-resolution image.

3. The lightweight mine image super-resolution reconstruction method based on deep feature calibration according to claim 2 is characterized by: In step 1, shallow feature extraction is performed on the input image through the shallow feature extraction layer, and the expression is: F0=Conv3(I LR ); Among them, I LR is the input image; F0 is the extracted shallow feature; Conv3(·) is the convolution operation with a kernel size of 3×3.

4. The lightweight mine image super-resolution reconstruction method based on deep feature calibration according to claim 2, characterized in that: The step 2.2 specifically includes the following steps: Step 2.2.

1. Use the output of the deep feature calibration module as the input feature, perform a 1×1 convolution operation on the input feature, and then perform a channel splitting operation to divide the input feature into two parts; Step 2.2.2: Use the parallel branches of maximum pooling and average pooling to input the features of step 2.2.1 into the maximum pooling branch and the average pooling branch respectively; Step 2.2.3: Use a convolutional layer with a kernel size of 1×1 and a Sigmoid activation function to multiply the outputs of the two branches with the corresponding inputs to perform feature high-frequency calibration. Step 2.2.4: The outputs of the two branches are fused together through a convolutional layer with a kernel size of 1×1 and Concat splicing to form the output of the attention feature.

5. The lightweight mine image super-resolution reconstruction method based on deep feature calibration according to claim 2, characterized in that: In step 2.3, the output of each deep feature calibration module and the output of the deep feature calibration attention module (HFFCA) are fused. Specifically, in the feature aggregation path, the feature output of each deep feature calibration module (HFFCB) is sent to the end of RAFG, and feature aggregation is performed. In the attention aggregation path, the attention output of each deep feature calibration module is directly sent to the deep feature calibration attention module to calibrate and select high-frequency information, and then transmitted to the end of RAFG for aggregation. Finally, the channel contrast attention CCA is used to adjust the feature information after the fusion of the feature aggregation path and the attention aggregation path, so that it focuses on the key and important channel content, and then added to the initial input.

6. The lightweight mine image super-resolution reconstruction method based on deep feature calibration according to claim 2, characterized in that: Step 3 aggregates the output of each residual attention feature module and fuses it with the shallow features; it is expressed as: Among them, F FC n ,F AC n represents the feature output and attention output of the nth residual attention feature module, Concat(·) represents the cascade on the channel dimension; Conv1(·) is a 1×1 convolution, F ff Represents the output after feature fusion.

7. The lightweight mine image super-resolution reconstruction method based on deep feature calibration according to claim 2, characterized in that: The step 4 inputs the fused features into the image reconstruction module layer for image reconstruction, and outputs the reconstructed super-resolution image I SR , expressed as: I SR =Conv3(H PixelShuffle (F ff +F0)) Among them, H PixelShuffle (·) is the sub-pixel convolution layer, I SR is the reconstructed super-resolution image.

Images