Lightweight image super-resolution reconstruction system and method for underground mobile equipment

Abstract

The present invention discloses a lightweight image super-resolution reconstruction system and method for underground mobile equipment, belonging to the field of image reconstruction technology; the system comprises: a shallow feature extraction module, a deep feature extraction module, a feature aggregation module, and an image reconstruction module; the deep feature extraction module comprises a high-frequency enhancement distillation block, which comprises a feature refinement module, a multi-scale information distillation module, and an information fusion module; the feature refinement module comprises three series-connected high-frequency enhancement residual blocks. The present invention adopts an information distillation architecture, employing a dual-path branch structure in the feature extraction backbone to perform channel segmentation on the input features, reducing the amount of computation, utilizing global feature extraction and high-frequency enhancement branches to enhance the network's ability to capture detailed information, and efficiently extracting deep information; a multi-scale information distillation method is employed to perform multi-scale high-frequency information distillation on the features refined by the feature extraction backbone, and a hybrid feature enhancement block is used to fuse the distilled multi-scale high-frequency information.

Description

Technical Field

[0001] The present invention belongs to the technical field of image reconstruction, and in particular relates to a lightweight image super-resolution reconstruction system and method for underground mobile equipment. Background Art

[0002] Mobile underground equipment is currently widely used in complex environments. For example, mining vehicles are often equipped with cameras to monitor their surroundings and avoid collisions. Some modern drilling rigs are equipped with image acquisition systems that can monitor the drilling process in real time and analyze rock formation structure. Remote-controlled exploration robots are equipped with high-resolution cameras and sensors to capture environmental images and data for underground exploration, inspection, and safety monitoring. Due to the complex and harsh underground environment, severe lack of illumination and reliance on limited lighting systems, images can suffer from uneven brightness and low clarity. Dust and moisture introduce noise, further degrading image quality. Improving the quality of images captured by these devices remains challenging, and image super-resolution (SR) reconstruction technology has emerged as a promising solution and is gaining increasing attention. SR is an image processing technique that uses algorithms to improve the quality of low-resolution images, resulting in higher-resolution visual effects. Currently, various deep learning-based image super-resolution reconstruction algorithms exist. For example, recent research utilizes conditional GANs (cGANs) to generate high-resolution images, enhancing image detail and realism through adversarial training. However, GAN models typically require significant computing resources, making real-time processing difficult on mobile devices. Models such as SRCNN, VDSR, and ESPCN can restore image details to a certain extent by increasing network depth and optimizing convolutional layer design. However, their high complexity limits their application in resource-constrained underground environments. To adapt to the needs of mobile devices, researchers have begun exploring lightweight network architectures such as MobileNet and ShuffleNet. These models improve real-time performance by reducing parameter count and computational complexity, but their performance is limited and they cannot adapt to the changing underground environment.

[0003] Although current similar technologies have made some progress in image super-resolution reconstruction, they still face challenges in downhole mobile device applications, such as high computational complexity and poor adaptability. More efficient and adaptable technical solutions are needed. First, in super-resolution reconstruction, high-frequency information is crucial for reconstructing image details. However, in downhole environments, detailed features are often weakened by factors such as low illumination and noise. This results in poor extraction of high-frequency information and insufficient detail restoration in reconstructed images. Existing technologies often require deeper and more complex network structures to compensate for the lack of detailed information, which significantly increases computational overhead and makes it difficult to meet real-time requirements. Second, downhole images contain a rich variety of information at different scales, and the detection of small devices and tiny targets requires this information. Existing super-resolution models mostly use fixed-scale feature extraction methods and lack effective multi-scale feature fusion mechanisms, making them inflexible in addressing the variable-scale detail in downhole environments. This results in the model often failing to simultaneously account for both overall structure and local details when processing the same image. Furthermore, the high computational complexity of these models makes them difficult to scale down. Downhole equipment places stringent requirements on the computational complexity and storage space of models. However, existing super-resolution models are typically deep neural networks with high computational complexity, high resource consumption, and high hardware performance requirements. This limits their application on mobile downhole equipment, especially in scenarios with high real-time processing demands, where rapid response and real-time monitoring cannot be guaranteed.

[0004] In summary, the existing algorithms have the following shortcomings:

[0005] 1. Most existing algorithms compensate for the lack of detailed information through deep and complex network structures, which requires a lot of computational overhead and lacks effective methods for extracting high-frequency information;

[0006] 2. Most existing algorithms use fixed-scale feature extraction methods, lack an effective fusion mechanism for multi-scale features, and cannot flexibly respond to the detailed information of varying scales in the underground environment.

[0007] 3. The existing model has high computational complexity and a large number of parameters, making it difficult to lightweight and unable to meet real-time requirements.

[0008] Therefore, how to solve the problem of detailed information with varying scales in underground environments and realize effective high-frequency information extraction. Summary of the Invention

[0009] The object of the present invention is to provide a lightweight image super-resolution reconstruction system and method for an underground mobile device, so as to solve the problems raised in the above background technology.

[0010] The object of the present invention is achieved as follows: a lightweight image super-resolution reconstruction system for underground mobile equipment, characterized in that: the reconstruction system includes a shallow feature extraction module, a deep feature extraction module, a feature aggregation module and an image reconstruction module;

[0011] The deep feature extraction module includes a high-frequency enhancement distillation block, which includes a feature refinement module, a multi-scale information distillation module and an information fusion module;

[0012] The feature refinement module includes three high-frequency enhancement residual blocks connected in series, which are used to extract and refine detail information layer by layer;

[0013] The multi-scale information distillation module includes four branches, including a first branch, a second branch, a third branch, and a fourth branch, each branch extracting high-frequency information through multi-scale sampling;

[0014] The information fusion module uses a hybrid feature enhancement block to perform information fusion.

[0015] Preferably, the high-frequency enhancement residual block includes global feature extraction, an upper high-frequency enhancement branch and a lower high-frequency enhancement branch, and the upper high-frequency enhancement branch includes a 3×3 convolution kernel, a 7×7 depth-separable convolution, a Sigmoid and a GELU activation function;

[0016] The lower high-frequency enhancement branch includes a maximum pooling layer, a 1×1 convolution kernel, and a GELU activation function. The maximum pooling layer extracts high-frequency information from the input features and enhances the high-frequency features through the 1×1 convolution kernel and the GELU activation function.

[0017] The outputs of the upper high-frequency enhancement branch and the lower high-frequency enhancement branch are connected through Concat and fed into a 1×1 convolution kernel to fuse the information, and finally a residual connection is introduced to fuse the information with the input.

[0018] Preferably, the four branches respectively use a 1×1 convolution kernel, the first branch inputs the original scale information, and the second branch, the third branch and the fourth branch all subtract the original input by downsampling and upsampling by different multiples to obtain multi-scale high-frequency information.

[0019] Preferably, the hybrid feature enhancement block HFEB includes Concat, C shuffle channels and 3×3 convolution kernel blocks. After the distillation of the four branches in the multi-scale information distillation module, it is connected through the Concat feature and then shuffled with the C shuffle channel to enhance the information exchange between different channels. Then, a 3×3 convolution kernel is used to further extract deep features.

[0020] A lightweight image super-resolution reconstruction method for underground mobile equipment is characterized by:

[0021] Step S1: obtaining an original image of an underground mobile device, and obtaining a low-resolution image using the original image;

[0022] Step S2: inputting the input low-resolution image into the reconstruction system;

[0023] Step S3: performing shallow feature extraction using a shallow feature extraction module;

[0024] Step S4: extract deep features using high-frequency enhancement distillation block;

[0025] Step S5: performing feature aggregation using a feature aggregation module;

[0026] Step S6: Reconstruct the image using a sub-pixel volume method;

[0027] Step S7: The reconstruction system outputs a super-resolution image.

[0028] Preferably, in step S3, shallow feature extraction is performed using a shallow feature extraction module, specifically:

[0029] The feature extraction module uses a 3×3 convolution to extract shallow features from the input image. The process is expressed as:

[0030] F0=h(I LR );

[0031] Among them, h is the shallow feature extraction function, and F0 is the extracted shallow feature.

[0032] Preferably, in step S4, high-frequency enhanced distillation blocks are used to extract deep features, specifically:

[0033] Step S4-1: The shallow features F0 extracted by the shallow feature extraction module are input into multiple recursively stacked high-frequency enhancement distillation blocks to gradually refine and extract deep features. The process is expressed as follows:

[0034] F k =H k (F k-1 ),k=1,...,m;

[0035] Among them, H k is the kth HFEDB feature extraction function; F k is the output feature of the kth HFEDB; F k-1 is the output feature of the k-1th HFEDB and the input feature of the kth HFEDB; when k=1, F0 is the extracted shallow feature;

[0036] Step S4-2: Input the original scale information through the first branch, use 1×1 convolution to halve the number of channels of the input feature, and use it as the feature after distillation of the first branch information;

[0037] Step S4-3: Through three series-connected high-frequency enhancement residual blocks (HFERBs), in each HFERB, the input features are respectively extracted and refined layer by layer through the upper high-frequency enhancement branch and the lower high-frequency enhancement branch;

[0038] Step S4-4: The information enhanced by the three high-frequency enhancement residual blocks is input into the second, third, and fourth branches respectively. Each branch is subtracted from the original input by downsampling and upsampling by different multiples to obtain multi-scale high-frequency information. Then, each branch uses 1×1 convolution to halve the number of channels of the input feature, which is used as the distilled features of the second, third, and fourth branches.

[0039] Step S4-5: The features after distillation of the four-way information are fused through the hybrid feature enhancement block HFEB. Let the input be F k-1 , the process is expressed as:

[0040] F Hi =H HFERB (F k-1 ),i=1,2,3;

[0041]

[0042] F k =H HFEB (F d );

[0043] Among them, H HFERB is the HFERB feature refinement function; F Hi For each HFERB output; F d is the feature after four-way information distillation; C1 is 1×1 convolution; It is a three-way downsampling operation, downsampling by 2 times, 4 times, and 8 times respectively; It is a three-way upsampling operation, upsampling by 2 times, 4 times, and 8 times respectively; H HFEB is the information fusion function.

[0044] Preferably, in step S5, feature aggregation is performed using a feature aggregation module, specifically:

[0045] The intermediate features output by the high-frequency enhancement distillation block are first aggregated by a 1×1 convolutional layer, then smoothed by a 3×3 convolutional layer, and finally added to the shallow features extracted in the first part. The process is expressed as:

[0046] F a =F0+H a(Concat(F out1 ,...,F outm ));

[0047] Among them, H a represents the feature aggregation function, F a is the aggregated feature, F out1 ,...,F outm It represents the intermediate features output from the 1st HFEDB to the mth HFEDB, and Concat represents the feature connection operation.

[0048] Preferably, in step S6, the sub-pixel volume method is used to reconstruct the image, specifically:

[0049] Assuming the target magnification is r, first use 3×3 convolution to change the number of channels of the input feature map to r 2 , and then r for each pixel of the feature image 2 The channels are rearranged into an r×r region, corresponding to an r×r sub-block in the high-resolution image, so that the size is H×W×r 2 The feature image is rearranged into a high-resolution image of rH×rW×1, and the reconstruction process is expressed as:

[0050] I SR =R(F a );

[0051] Among them, R represents the reconstruction function, I SR is the output super-resolution image of the network.

[0052] Compared with the prior art, the present invention has the following improvements and advantages:

[0053] 1. Adopting an information distillation architecture, a dual-branch structure is adopted in the feature extraction backbone to perform channel segmentation on the input features, reducing the amount of computation. Global feature extraction and high-frequency enhancement branches are then used to improve the network's ability to capture detailed information and efficiently extract deep information.

[0054] 2. A multi-scale information distillation method is used to perform multi-scale high-frequency information distillation on the features after feature extraction backbone refinement, and the mixed feature enhancement block is used to fuse the distilled multi-scale high-frequency information.

[0055] 3. Construct a high-frequency enhancement distillation block. By connecting high-frequency enhancement residual blocks in series and utilizing a multi-scale high-frequency information distillation method, we optimize the high-frequency information extraction method, efficiently extract and fuse multi-scale information, and ensure the lightweight model while improving network performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a structural diagram of the reconstruction system of the present invention.

[0057] Figure 2 Flowchart of the method of the present invention.

[0058] Figure 3 This is a structural diagram of the high frequency enhancement distillation module HFEDB in the present invention.

[0059] Figure 4 This is a structural diagram of the high frequency enhanced residual block HFERB in the present invention.

[0060] Figure 5 This is a structural diagram of the hybrid feature enhancement block HFEB in the present invention. DETAILED DESCRIPTION

[0061] The present invention is further summarized below with reference to the accompanying drawings.

[0062] like Figure 1 As shown, a lightweight image super-resolution reconstruction system for underground mobile equipment includes a shallow feature extraction module, a deep feature extraction module, a feature aggregation module, and an image reconstruction module;

[0063] The shallow feature extraction module extracts shallow features from the input image through a 3×3 convolution. The process is expressed as:

[0064] F0=h(I LR );

[0065] Among them, h is the shallow feature extraction function, and F0 is the extracted shallow feature.

[0066] like Figure 3 As shown, the deep feature extraction module includes a high-frequency enhancement distillation block, which includes a feature refinement module, a multi-scale information distillation module, and an information fusion module;

[0067] The feature refinement module consists of three serially connected high-frequency enhancement residual blocks, which are used to extract and refine detail information layer by layer; Figure 4 As shown in the figure, the high-frequency enhancement residual block includes global feature extraction, an upper high-frequency enhancement branch, and a lower high-frequency enhancement branch. The upper high-frequency enhancement branch includes a 3×3 convolution kernel, a 7×7 depth-separable convolution, a Sigmoid and a GELU activation function; the 3×3 convolution is used to extract local features, while the 7×7 depth-separable convolution can expand the receptive field while ensuring computational efficiency. The Sigmoid function generates weights that reflect the global importance of pixels, which are multiplied pixel by pixel with the extracted local features to enhance features that are beneficial to image restoration and are output through the GELU activation function.

[0068] The lower high-frequency enhancement branch includes a maximum pooling layer, a 1×1 convolution kernel, and a GELU activation function. The maximum pooling layer extracts high-frequency information from the input features and enhances the high-frequency features through a 1×1 convolution kernel and a GELU activation function. The outputs of the upper and lower high-frequency enhancement branches are connected through Concat and fed into a 1×1 convolution kernel to fuse the information. Finally, a residual connection is introduced to fuse the information with the input information.

[0069] The multi-scale information distillation module includes four branches, namely the first branch, the second branch, the third branch and the fourth branch. Each branch extracts high-frequency information through multi-scale sampling; the information fusion module uses a hybrid feature enhancement block for information fusion.

[0070] The four branches use 1×1 convolution kernels respectively. The first branch inputs the original scale information, and the second, third and fourth branches subtract the original input by downsampling and upsampling with different multiples to obtain multi-scale high-frequency information.

[0071] The information fusion module uses a hybrid feature enhancement block to perform information fusion, such as Figure 5 As shown in the figure, the hybrid feature enhancement block HFEB includes Concat, C shuffle channels and 3×3 convolution kernel blocks. After the distillation of the four branches in the multi-scale information distillation module, it is connected through Concat features and then shuffled through C shuffle channels to enhance the information exchange between different channels. Then, a 3×3 convolution kernel is used to further extract deep features.

[0072] like Figure 2 As shown, a lightweight image super-resolution reconstruction method for an underground mobile device includes the following steps:

[0073] Step S1: obtaining an original image of an underground mobile device, and obtaining a low-resolution image using the original image;

[0074] Step S2: inputting the input low-resolution image into the reconstruction system;

[0075] Step S3: performing shallow feature extraction using a shallow feature extraction module;

[0076] Step S4: extract deep features using high-frequency enhancement distillation block;

[0077] The high-frequency enhancement distillation block is used to extract deep features, specifically:

[0078] Step S4-1: The shallow features F0 extracted by the shallow feature extraction module are input into multiple recursively stacked high-frequency enhancement distillation blocks to gradually refine and extract deep features. The process is expressed as follows:

[0079] F k =Hk (F k-1 ),k=1,...,m;

[0080] Among them, H k is the kth HFEDB feature extraction function; F k is the output feature of the kth HFEDB; F k-1 is the output feature of the k-1th HFEDB and the input feature of the kth HFEDB; when k=1, F0 is the extracted shallow feature;

[0081] Step S4-2: Input the original scale information through the first branch, use 1×1 convolution to halve the number of channels of the input feature, and use it as the feature after distillation of the first branch information;

[0082] Step S4-3: Through three series-connected high-frequency enhancement residual blocks (HFERBs), in each HFERB, the input features are respectively extracted and refined layer by layer through the upper high-frequency enhancement branch and the lower high-frequency enhancement branch;

[0083] Step S4-4: The information enhanced by the three high-frequency enhancement residual blocks is input into the second, third, and fourth branches respectively. Each branch is subtracted from the original input by downsampling and upsampling by different multiples to obtain multi-scale high-frequency information. Then, each branch uses 1×1 convolution to halve the number of channels of the input feature, which is used as the distilled features of the second, third, and fourth branches.

[0084] Step S4-5: The features after distillation of the four-way information are fused through the hybrid feature enhancement block HFEB. Let the input be F k-1 , the process is expressed as:

[0085] F Hi =H HFERB (F k-1 ),i=1,2,3;

[0086]

[0087] F k =H HFEB (F d );

[0088] Among them, H HFERB is the HFERB feature refinement function; F Hi For each HFERB output; F d is the feature after four-way information distillation; C1 is 1×1 convolution; It is a three-way downsampling operation, downsampling by 2 times, 4 times, and 8 times respectively; It is a three-way upsampling operation, upsampling by 2 times, 4 times, and 8 times respectively; H HFEB is the information fusion function.

[0089] Step S5: performing feature aggregation using a feature aggregation module;

[0090] Use the feature aggregation module to perform feature aggregation, specifically:

[0091] The intermediate features output by the high-frequency enhancement distillation block are first aggregated by a 1×1 convolutional layer, then smoothed by a 3×3 convolutional layer, and finally added to the shallow features extracted in the first part. The process is expressed as:

[0092] F a =F0+H a (Concat(F out1 ,...,F outm ));

[0093] Among them, H a represents the feature aggregation function, F a is the aggregated feature, F out1 ,...,F outm It represents the intermediate features output from the 1st HFEDB to the mth HFEDB, and Concat represents the feature connection operation.

[0094] Step S6: Reconstruct the image using a sub-pixel volume method;

[0095] The sub-pixel volume method is used for image reconstruction, specifically:

[0096] Assuming the target magnification is r, first use 3×3 convolution to change the number of channels of the input feature map to r 2 , and then r for each pixel of the feature image 2 The channels are rearranged into an r×r region, corresponding to an r×r sub-block in the high-resolution image, so that the size is H×W×r 2 The feature image is rearranged into a high-resolution image of rH×rW×1, and the reconstruction process is expressed as:

[0097] I SR =R(F a );

[0098] Among them, R represents the reconstruction function, I SR is the output super-resolution image of the network.

[0099] Step S7: The reconstruction system outputs a super-resolution image.

[0100] This invention proposes a high-frequency enhancement residual block with dual branches, which improves the network's ability to capture detailed information. Compared with traditional deep and complex networks, this efficient design can reduce computational overhead while maintaining the integrity of detailed information.

[0101] Most existing algorithms use fixed-scale feature extraction methods and lack effective fusion mechanisms for multi-scale features. This makes them inflexible when dealing with the variable-scale details of downhole environments, which in turn affects reconstruction quality. This paper proposes a multi-scale sampling high-frequency information distillation method. By performing multi-scale high-frequency information distillation on the refined features and combining it with a hybrid feature enhancement block for effective fusion, this method enhances the model's ability to handle variable detail information in downhole environments, thereby improving reconstruction quality.

[0102] This invention optimizes the extraction of high-frequency information by cascading high-frequency enhancement residual blocks with a multi-scale high-frequency information distillation method. This design not only improves network performance but also fully considers the lightweight model, ensuring real-time performance on mobile underground equipment and meeting the stringent requirements of computational complexity and parameter count.

[0103] 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 image super-resolution reconstruction system for underground mobile equipment, characterized by: The reconstruction system includes a shallow feature extraction module, a deep feature extraction module, a feature aggregation module and an image reconstruction module; The deep feature extraction module includes a high-frequency enhancement distillation block, which includes a feature refinement module, a multi-scale information distillation module and an information fusion module; The feature refinement module includes three high-frequency enhancement residual blocks connected in series, which are used to extract and refine detail information layer by layer; The multi-scale information distillation module includes four branches, including a first branch, a second branch, a third branch, and a fourth branch, each branch extracting high-frequency information through multi-scale sampling; The high-frequency enhancement residual block includes global feature extraction, an upper high-frequency enhancement branch, and a lower high-frequency enhancement branch. The upper high-frequency enhancement branch includes a 3×3 convolution kernel, a 7×7 depth-separable convolution, a Sigmoid and a GELU activation function. The lower high-frequency enhancement branch includes a maximum pooling layer, a 1×1 convolution kernel, and a GELU activation function. The maximum pooling layer extracts high-frequency information from the input features and enhances the high-frequency features through the 1×1 convolution kernel and the GELU activation function. The outputs of the upper high-frequency enhancement branch and the lower high-frequency enhancement branch are connected through Concat and fed into a 1×1 convolution kernel to fuse the information, and finally a residual connection is introduced to fuse the information with the input; The information fusion module uses a hybrid feature enhancement block to perform information fusion.

2. The lightweight image super-resolution reconstruction system for underground mobile equipment according to claim 1, characterized in that: The four branches respectively use a 1×1 convolution kernel. The first branch inputs the original scale information, and the second branch, the third branch, and the fourth branch all subtract the original input by downsampling and upsampling by different multiples to obtain multi-scale high-frequency information.

3. The lightweight image super-resolution reconstruction system for underground mobile equipment according to claim 1, characterized in that: The hybrid feature enhancement block (HFEB) includes Concat, C shuffle channels, and a 3×3 convolution kernel block. After distillation of four branches in the multi-scale information distillation module, it is connected through Concat features and then shuffled through C shuffle channels to enhance information exchange between different channels. A 3×3 convolution kernel is then used to further extract deep features.

4. The method for generating and reconstructing a lightweight image super-resolution reconstruction system for an underground mobile device according to any one of claims 1 to 3, characterized in that: Step S1: obtaining an original image of an underground mobile device, and obtaining a low-resolution image using the original image; Step S2: inputting the input low-resolution image into the reconstruction system; Step S3: performing shallow feature extraction using a shallow feature extraction module; Step S4: extract deep features using high-frequency enhancement distillation block; Step S5: performing feature aggregation using a feature aggregation module; Step S6: Reconstruct the image using a sub-pixel volume method; Step S7: The reconstruction system outputs a super-resolution image.

5. The lightweight image super-resolution reconstruction method for underground mobile equipment according to claim 4, characterized in that: In step S3, shallow feature extraction is performed using a shallow feature extraction module, specifically: The shallow feature extraction module uses a 3×3 convolution to extract shallow features from the input image. The process is expressed as: F0=h(I LR ); Among them, h is the shallow feature extraction function, F0 is the extracted shallow feature, and LR is the input image.

6. The lightweight image super-resolution reconstruction method for underground mobile equipment according to claim 4, characterized in that: In step S4, the high-frequency enhancement distillation block is used to extract deep features, specifically: Step S4-1: The shallow features F0 extracted by the shallow feature extraction module are input into multiple recursively stacked high-frequency enhancement distillation blocks to gradually refine and extract deep features. The process is expressed as follows: F k =H k (F k-1 ),k=1,...,m; Among them, H k is the kth HFEDB feature extraction function; F k is the output feature of the kth HFEDB; F k-1 is the output feature of the k-1th HFEDB and the input feature of the kth HFEDB; when k = 1, F0 is the extracted shallow feature, and m is the number of high-frequency enhancement distillation blocks; Step S4-2: Input the original scale information through the first branch, use 1×1 convolution to halve the number of channels of the input feature, and use it as the feature after distillation of the first branch information; Step S4-3: Through three series-connected high-frequency enhancement residual blocks (HFERBs), in each HFERB, the input features are respectively extracted and refined layer by layer through the upper high-frequency enhancement branch and the lower high-frequency enhancement branch; Step S4-4: The information enhanced by the three high-frequency enhancement residual blocks is input into the second, third, and fourth branches respectively. Each branch is subtracted from the original input by downsampling and upsampling by different multiples to obtain multi-scale high-frequency information. Then, each branch uses 1×1 convolution to halve the number of channels of the input feature, which is used as the distilled features of the second, third, and fourth branches. Step S4-5: The features after distillation of the four-way information are fused through the hybrid feature enhancement block HFEB. Let the input be F k-1 , the process is expressed as: F Hi =H HFERB (F k-1 ),i=1,2,3; F k =H HFEB (F d ); Among them, H HFERB is the HFERB feature refinement function; F Hi For each HFERB output; F d is the feature after four-way information distillation; C1 is 1×1 convolution; It is a three-way downsampling operation, downsampling by 2 times, 4 times, and 8 times respectively; It is a three-way upsampling operation, upsampling by 2 times, 4 times, and 8 times respectively; H HFEB is the information fusion function.

7. The lightweight image super-resolution reconstruction method for underground mobile equipment according to claim 4, characterized in that: In step S5, feature aggregation is performed using a feature aggregation module, specifically: The intermediate features output by the high-frequency enhancement distillation block are first aggregated by a 1×1 convolutional layer, then smoothed by a 3×3 convolutional layer, and finally added to the shallow features extracted in the first part. The process is expressed as: F a =F0+H a (Concat(F out1 ,...,F outm )); Among them, H a represents the feature aggregation function, F a is the aggregated feature, F out1 ,...,F outm It represents the intermediate features output from the 1st HFEDB to the mth HFEDB, and Concat represents the feature connection operation.

8. The lightweight image super-resolution reconstruction method for underground mobile equipment according to claim 4, characterized in that: In step S6, the sub-pixel volume method is used to reconstruct the image, specifically: Assuming the target magnification is r, first use 3×3 convolution to change the number of channels of the input feature map to r 2 , and then r for each pixel of the feature image 2 The channels are rearranged into an r×r region, corresponding to an r×r sub-block in the high-resolution image, so that the size is H×W×r 2 The feature image is rearranged into a high-resolution image of rH×rW×1, and the reconstruction process is expressed as: I SR =R(F a ); Among them, R represents the reconstruction function, I SR is the output super-resolution image of the network.

Images