A method for super-resolution reconstruction of industrial images based on a physical consistency self-supervised mechanism
By constructing a lightweight neural network based on a physical consistency self-supervised mechanism, simulating the degradation process of industrial images, and combining training with multiple loss functions, the problem of insufficient generalization ability of traditional methods in industrial image super-resolution reconstruction is solved, and high-precision image enhancement effect is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2025-07-21
- Publication Date
- 2026-07-10
AI Technical Summary
In industrial image super-resolution reconstruction, existing technologies suffer from poor generalization ability of traditional supervised methods, difficulty in obtaining paired training samples, and failure of self-supervised methods to effectively simulate the physical degradation process, resulting in poor image enhancement effects.
A lightweight neural network based on a physical consistency self-supervised mechanism is constructed. By simulating the non-uniform degradation process and optical imaging blur of industrial scene images, the basic reconstruction loss, structural edge perception loss and optical blur consistency loss are combined for training on unlabeled data.
It significantly improves image edge restoration capabilities and detail reconstruction accuracy, making it suitable for industrial image enhancement tasks that do not require high-quality paired data, and applicable to surface mount equipment and automated optical inspection systems.
Smart Images
Figure CN120876229B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of industrial automation and machine vision, specifically relating to an industrial image super-resolution reconstruction method based on a physical consistency self-supervised mechanism. Background Technology
[0002] In industrial automation and intelligent manufacturing systems, machine vision has become an important means of achieving precision recognition and measurement. However, in actual industrial scenarios, due to the influence of sensor noise, optical imaging blurring, ambient light interference, and motion jitter, the acquired images often suffer from severe degradation problems, such as low resolution, blurred edges, and loss of structural details. This severely restricts subsequent image recognition and parameter extraction tasks.
[0003] Most mainstream image super-resolution reconstruction methods currently rely on supervised training with paired high- and low-resolution images. While typical supervised methods perform excellently on ideal datasets, in real industrial images, the degradation process is complex and variable, and cannot be accurately modeled. Traditional supervised methods have poor generalization ability and are difficult to obtain paired training samples, thus limiting their practical applications. Although some self-supervised methods can be trained independently of paired data, they still do not explicitly model the physical degradation process, nor are they optimized for structural restoration tasks, making it difficult to meet the needs of high-precision industrial image enhancement. Summary of the Invention
[0004] The problem this invention aims to solve is to obtain high-resolution industrial images, and proposes an industrial image super-resolution reconstruction method based on a physical consistency self-supervised mechanism.
[0005] To achieve the above objectives, the present invention provides the following technical solution:
[0006] An industrial image super-resolution reconstruction method based on a physical consistency self-supervised mechanism includes the following steps:
[0007] S1. A method for generating non-uniformly degraded images of industrial scenes is constructed, including acquiring industrial scene images and preprocessing them, then dividing the preprocessed industrial scene images into regions, adding composite noise perturbation according to the region brightness, and generating non-uniformly degraded images of industrial scene images.
[0008] S2. Construct an imaging blur degradation model based on the Airy Disk point spread function to simulate the blur degradation process of industrial scene images, so as to calculate the physical consistency loss in lightweight neural network training;
[0009] S3. Construct a lightweight neural network consisting of multi-layer convolutional modules and pixel recombination modules;
[0010] S4. Construct a composite loss function that includes basic reconstruction loss, structural edge perception loss and optical blur consistency loss. After preprocessing the industrial scene images, input them into the lightweight neural network constructed in step S3 for training to obtain the trained lightweight neural network.
[0011] S5. Acquire industrial scene images, then process them using a multi-view spatial transformation and integration mechanism, and input them into the lightweight neural network trained in step S4 to output an enhanced image of the industrial scene.
[0012] Furthermore, the specific implementation method of step S1 includes the following steps:
[0013] S1.1. Acquire industrial scene images and perform preprocessing, including channel expansion and median filtering, to obtain preprocessed industrial scene images. ;
[0014] S1.2. Then, the preprocessed industrial scene image obtained in step S1.1 is divided into... Given a fixed-size, non-overlapping sub-region, the i-th sub-region is obtained as follows: ;
[0015] S1.3. Calculate the average gray value of each sub-region obtained in step S1.2, and obtain the average gray value of the i-th sub-region. ;
[0016] S1.4. Based on the average gray value of the sub-region obtained in step S1.3, the Poisson noise intensity coefficient and Gaussian noise intensity coefficient of each region are adaptively adjusted to obtain the expression for the noise generation process:
[0017]
[0018] in, Generate an industrial scene image for the noise in the i-th sub-region. Represents the Poisson process. Indicates Gaussian noise. Let be the Gaussian noise intensity coefficient of the i-th sub-region. Let be the Poisson noise intensity coefficient of the i-th sub-region. , ;
[0019] Then, a non-uniform degradation image of the industrial scene is obtained. Represented as:
[0020] .
[0021] Furthermore, the specific implementation method of step S2 includes the following steps:
[0022] S2.1. Set up the sample The observed image formed after passing through the optical imaging system The noise term is represented by the convolution process of the sample and the point spread function of the optical imaging system, and its expression is:
[0023]
[0024] in, The point spread function represents the optical imaging system. This refers to spatial noise during the imaging process. Represents a two-dimensional convolution operation;
[0025] Set point diffusion function It is expressed in the radial form of the point spread function. The distance from the midpoint of the image plane to the optical axis is expressed as:
[0026] ;
[0027] S2.2. Constructing an imaging blur degradation model based on the Airy Disk point spread function The Airy Disk PSF model based on the ideal circular aperture diffraction theory is introduced to model the imaging blur, and the expression is obtained as follows:
[0028]
[0029] in, It is a first-order Bessel function. The normalized spatial frequency parameter;
[0030]
[0031]
[0032] in, For imaging wavelength, For the summation index in the expansion of an infinite series, is the gamma function, used to generalize factorial, and NA is the numerical aperture.
[0033] Furthermore, step S3 constructs a lightweight neural network consisting of a multi-layer convolutional module and a pixel reconstruction module, which consists of three 3×3 convolutional layers and a pixel reconstruction upsampling module.
[0034] Furthermore, the specific implementation method for constructing the composite loss function, which includes the basic reconstruction loss, the structural edge perception loss, and the optical blur consistency loss, in step S4 includes the following steps:
[0035] S4.1. Construct the basic reconstruction loss using the standard loss function. , used to measure the pixel-level difference between the reconstructed image and the reference image, is defined as:
[0036]
[0037] in, Predict images for the network, For preprocessing images;
[0038]
[0039] in, Indicates preprocessed image The sampling factor is The network training input image is formed by downsampling and then adding noise through a non-uniform degradation image generation method for industrial scene images;
[0040] S4.2. Construct a structural edge-aware loss function using edge masks. The expression for measuring the gradient change in an image is:
[0041]
[0042] in, and These are the first-order differences in the horizontal and vertical directions of the image, respectively. It is the numerical stability constant;
[0043] Structural edge sensing loss Use edge masking The reconstruction error is weighted to guide the network to focus on structurally sensitive regions in the image. The expression is as follows:
[0044] ;
[0045] S4.3. Constructing an optical fuzziness consistency loss Network predicts images An auxiliary pseudo-supervisory image is generated by using an imaging blur degradation model based on the Airy Disk point spread function and adding regional noise through a non-uniform degradation image generation method for industrial scene images. The expression is:
[0046] ;
[0047] S4.4. Construct a composite loss function that includes basic reconstruction loss, structural edge perception loss, and optical blur consistency loss. The resulting expression is:
[0048]
[0049] in, The weight parameters are those for the structure edge-aware loss. The weighting parameters for optical blur consistency loss are: , .
[0050] Furthermore, in step S4, during the training of the lightweight neural network, the target image is the preprocessed image. The image generated by performing channel averaging, with the input image being... The reconstruction results output by the lightweight neural network are averaged across channels and then optimized using a composite loss function. After training, the lightweight neural network obtains the reconstruction results from the self-supervised low-resolution image. ,in This represents a trained lightweight neural network. This represents images of an industrial scene that have been collected.
[0051] Furthermore, the industrial scene images acquired in step S5 Through a set of geometric transformation operations Then, the results are fed into a pre-trained lightweight neural network to obtain preliminary reconstruction results, followed by inverse transformation. Restored to the original space, the final output image is obtained. The expression is:
[0052] .
[0053] Furthermore, in step S5, a set of geometric transformation operations are horizontal flip, vertical flip, 90-degree, 180-degree, and 270-degree rotation.
[0054] The beneficial effects of this invention are:
[0055] This invention presents an industrial image super-resolution reconstruction method based on a physically consistent self-supervised mechanism. By introducing a region-aware noise perturbation mechanism and optical point spread function modeling, it approximates the simulation of real non-uniform degradation and imaging blurring processes in industrial images, enabling the network to effectively learn image structural information without the need for labeled data. Combining structure-aware loss and physically consistent loss significantly improves image edge restoration capability and detail reconstruction accuracy. The inference stage employs a multi-view transformation ensemble strategy, effectively enhancing the stability and consistency of the output image. Compared to existing methods, this invention does not rely on high-quality paired data and has advantages such as strong adaptability in the training process, lightweight model structure, and high inference efficiency. It is suitable for offline image enhancement tasks in industrial vision systems such as surface mount equipment and automated optical inspection (AOI). Attached Figure Description
[0056] Figure 1 This is a flowchart of an industrial image super-resolution reconstruction method based on a physical consistency self-supervised mechanism as described in this invention.
[0057] Figure 2 This is a flowchart of the self-supervised training process described in this invention;
[0058] Figure 3 This is the self-integrated reasoning flowchart described in this invention;
[0059] Figure 4 This is a schematic diagram of the backbone network structure described in this invention. Detailed Implementation
[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described specific embodiments are merely a part of the embodiments of the invention, and not all of them. The components of the specific embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations, and the invention may also have other embodiments.
[0061] Therefore, the following detailed description of specific embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected specific embodiments of the invention. All other specific embodiments obtained by those skilled in the art based on these specific embodiments without inventive effort are within the scope of protection of this invention.
[0062] To further understand the invention's content, features, and effects, the following specific embodiments are provided, along with accompanying drawings. Figure 1 -Appendix Figure 4 Detailed explanation is as follows:
[0063] Example 1:
[0064] An industrial image super-resolution reconstruction method based on a physical consistency self-supervised mechanism includes the following steps:
[0065] S1. A method for generating non-uniformly degraded images of industrial scenes is constructed, including acquiring industrial scene images and preprocessing them, then dividing the preprocessed industrial scene images into regions, adding composite noise perturbation according to the region brightness, and generating non-uniformly degraded images of industrial scene images.
[0066] Furthermore, the specific implementation method of step S1 includes the following steps:
[0067] S1.1. Acquire industrial scene images and perform preprocessing, including channel expansion and median filtering, to obtain preprocessed industrial scene images. ;
[0068] S1.2. Then, the preprocessed industrial scene image obtained in step S1.1 is divided into... Given a fixed-size, non-overlapping sub-region, the i-th sub-region is obtained as follows: ;
[0069] S1.3. Calculate the average gray value of each sub-region obtained in step S1.2, and obtain the average gray value of the i-th sub-region. ;
[0070] S1.4. Based on the average gray value of the sub-region obtained in step S1.3, the Poisson noise intensity coefficient and Gaussian noise intensity coefficient of each region are adaptively adjusted to obtain the expression for the noise generation process:
[0071]
[0072] in, Generate an industrial scene image for the noise in the i-th sub-region. Represents the Poisson process. Indicates Gaussian noise. Let be the Gaussian noise intensity coefficient of the i-th sub-region. Let be the Poisson noise intensity coefficient of the i-th sub-region. , ;
[0073] Then, a non-uniform degradation image of the industrial scene is obtained. Represented as:
[0074] .
[0075] S2. Construct an imaging blur degradation model based on the Airy Disk point spread function to simulate the blur degradation process of industrial scene images, so as to calculate the physical consistency loss in lightweight neural network training;
[0076] Furthermore, the specific implementation method of step S2 includes the following steps:
[0077] S2.1. Set up the sample The observed image formed after passing through the optical imaging system The noise term is represented by the convolution process of the sample and the point spread function of the optical imaging system, and its expression is:
[0078]
[0079] in, The point spread function represents the optical imaging system. This refers to spatial noise during the imaging process. Represents a two-dimensional convolution operation;
[0080] Set point diffusion function It is expressed in the radial form of the point spread function. The distance from the midpoint of the image plane to the optical axis is expressed as:
[0081] ;
[0082] S2.2. Constructing an imaging blur degradation model based on the Airy Disk point spread function The Airy Disk PSF model based on the ideal circular aperture diffraction theory is introduced to model the imaging blur, and the expression is obtained as follows:
[0083]
[0084] in, It is a first-order Bessel function. The normalized spatial frequency parameter;
[0085]
[0086]
[0087] in, For imaging wavelength, For the summation index in the expansion of an infinite series, is the gamma function, used to generalize factorial, and NA is the numerical aperture.
[0088] S3. Construct a lightweight neural network consisting of multi-layer convolutional modules and pixel recombination modules;
[0089] Furthermore, step S3 constructs a lightweight neural network consisting of a multi-layer convolutional module and a pixel recombination module, which consists of three 3×3 convolutional layers and a pixel recombination upsampling module.
[0090] The connection relationship is as follows: input layer → first 3×3 convolutional layer → first ReLU activation layer → second 3×3 convolutional layer → second ReLU activation layer → third 3×3 convolutional layer → pixel shuffle layer → output layer.
[0091] S4. Construct a composite loss function that includes basic reconstruction loss, structural edge perception loss and optical blur consistency loss. After preprocessing the industrial scene images, input them into the lightweight neural network constructed in step S3 for training to obtain the trained lightweight neural network.
[0092] Furthermore, the specific implementation method for constructing the composite loss function, which includes the basic reconstruction loss, the structural edge perception loss, and the optical blur consistency loss, in step S4 includes the following steps:
[0093] S4.1. Construct the basic reconstruction loss using the standard loss function. , used to measure the pixel-level difference between the reconstructed image and the reference image, is defined as:
[0094]
[0095] in, Predict images for the network, For preprocessing images;
[0096]
[0097] in, Indicates preprocessed image The sampling factor is The network training input image is formed by downsampling and then adding noise through a non-uniform degradation image generation method for industrial scene images;
[0098] To enhance the model's responsiveness to key structural regions (such as edge contours, interface seams, and minor defects), an edge-aware loss function based on image gradients is introduced to improve the network's ability to learn local structural information.
[0099] S4.2. Construct a structural edge-aware loss function using edge masks. The expression for measuring the gradient change in an image is:
[0100]
[0101] in, and These are the first-order differences in the horizontal and vertical directions of the image, respectively. It is a numerical stability constant; the numerical stability constant is used to avoid the problem of the denominator being zero in gradient normalization.
[0102] Structural edge sensing loss Use edge masking The reconstruction error is weighted to guide the network to focus on structurally sensitive regions in the image. The expression is as follows:
[0103] ;
[0104] S4.3. Constructing an optical fuzziness consistency loss Network predicts images An auxiliary pseudo-supervisory image is generated by using an imaging blur degradation model based on the Airy Disk point spread function and adding regional noise through a non-uniform degradation image generation method for industrial scene images. The expression is:
[0105] ;
[0106] S4.4. Construct a composite loss function that includes basic reconstruction loss, structural edge perception loss, and optical blur consistency loss. The resulting expression is:
[0107]
[0108] in, The weight parameters are those for the structure edge-aware loss. The weighting parameters for optical blur consistency loss are: , .
[0109] Furthermore, in step S4, during the training of the lightweight neural network, the target image is the preprocessed image. The image generated by performing channel averaging, with the input image being... The reconstruction results output by the lightweight neural network are averaged across channels and then optimized using a composite loss function. After training, the lightweight neural network obtains the reconstruction results from the self-supervised low-resolution image. ,in This represents a trained lightweight neural network. This represents images of an industrial scene that have been collected.
[0110] S5. Acquire industrial scene images, then process them using a multi-view spatial transformation and integration mechanism, and input them into the lightweight neural network trained in step S4 to output an enhanced image of the industrial scene.
[0111] Furthermore, the industrial scene images acquired in step S5 Through a set of geometric transformation operations Then, the results are fed into a pre-trained lightweight neural network to obtain preliminary reconstruction results, followed by inverse transformation. Restored to the original space, the final output image is obtained. The expression is:
[0112] .
[0113] Furthermore, in step S5, a set of geometric transformation operations are horizontal flip, vertical flip, 90-degree, 180-degree, and 270-degree rotation.
[0114] To verify the effectiveness of the proposed method, the reconstruction performance of multiple algorithms was compared at ×2 and ×4 magnifications based on three typical electronic components (box-type components, pin-type components, and ball grid components). The results of quantitative evaluation of different algorithms for the three types of electronic components are shown in Table 1.
[0115] Table 1
[0116]
[0117] Table 1 shows that the method of this embodiment achieves excellent PSNR and SSIM indices at all types and magnifications, especially maintaining the structural reduction advantage under low magnification ×4 conditions.
[0118] Furthermore, to evaluate the image's ability to support subsequent parameter extraction tasks, three indicators—parameter measurement error RMSE, number of connected components, and contour length—were used for quantitative analysis. The quantitative comparison results of parameter extraction and binarization structure quality obtained by the method in this embodiment are shown in Table 2.
[0119] Table 2
[0120]
[0121] The experimental results in Table 2 show that the method in this embodiment retains structural information while avoiding the effects of image artifacts and excessive smoothing, and performs stably in connected component and contour measurements.
[0122] Finally, the model complexity and inference efficiency of various methods were compared under the same image size (500×500), and the results are shown in Table 3. Table 3 is a comparison table of model complexity and inference efficiency of the method in this embodiment;
[0123] Table 3
[0124]
[0125] As shown in Table 3, this embodiment uses only 0.0797M parameters and 0.0385G of computation, while maintaining accuracy and controlling the inference time to 6.47 seconds, which is far superior to methods such as ZSSR, demonstrating its industrial deployment friendliness. In summary, the method in this embodiment provides an industrial image super-resolution method that combines regional noise perturbation, physical fuzzy modeling, and structure-aware optimization. It features label-free operation, high accuracy, and deployability, and has good application value in real-world electronic manufacturing scenarios.
[0126] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0127] Although this application has been described above with reference to specific embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of this application. In particular, as long as there is no structural conflict, the features in the specific embodiments disclosed in this application can be combined with each other in any way. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, this application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.
Claims
1. A method for super-resolution reconstruction of industrial images based on a physical consistency self-supervised mechanism, characterized in that, Includes the following steps: S1. A method for generating non-uniformly degraded images of industrial scenes is constructed, including acquiring industrial scene images and preprocessing them, then dividing the preprocessed industrial scene images into regions, adding composite noise perturbation according to the region brightness, and generating non-uniformly degraded images of industrial scene images. S2. Construct an imaging blur degradation model based on the Airy Disk point spread function to simulate the blur degradation process of industrial scene images, so as to calculate the physical consistency loss in lightweight neural network training; The specific implementation method of step S2 includes the following steps: S2.
1. Setting up samples The observed image formed after passing through the optical imaging system The noise term is represented by the convolution process of the sample and the point spread function of the optical imaging system, and its expression is: ； in, The point spread function represents the optical imaging system. This refers to spatial noise during the imaging process. Represents a two-dimensional convolution operation; Set point diffusion function It is expressed in the radial form of the point spread function. The distance from the midpoint of the image plane to the optical axis is expressed as: ； S2.
2. Constructing an imaging blur degradation model based on the Airy Disk point spread function The Airy Disk PSF model based on the ideal circular aperture diffraction theory is introduced to model the imaging blur, and the expression is obtained as follows: ； in, It is a first-order Bessel function. The normalized spatial frequency parameter; ； ； in, For imaging wavelength, For the summation index in the expansion of an infinite series, NA is the gamma function, used to generalize factorials; NA is the numerical aperture. S3. Construct a lightweight neural network consisting of multi-layer convolutional modules and pixel recombination modules; S4. Construct a composite loss function that includes basic reconstruction loss, structural edge perception loss and optical blur consistency loss. After preprocessing the industrial scene images, input them into the lightweight neural network constructed in step S3 for training to obtain the trained lightweight neural network. The specific implementation method for constructing the composite loss function, which includes the basic reconstruction loss, the structural edge perception loss, and the optical blur consistency loss, in step S4 includes the following steps: S4.
1. Construct the basic reconstruction loss using the standard loss function. , used to measure the pixel-level difference between the reconstructed image and the reference image, is defined as: ； in, Predict images for the network. For preprocessing images; ； in, Indicates preprocessed image The sampling factor is The network training input image is formed by downsampling and then adding noise through a non-uniform degradation image generation method for industrial scene images; S4.
2. Construct a structural edge-aware loss function using edge masks. The expression for measuring the gradient change in an image is: ； in, and These are the first-order differences in the horizontal and vertical directions of the image, respectively. It is the numerical stability constant; Structural edge sensing loss Use edge masking The reconstruction error is weighted to guide the network to focus on structurally sensitive regions in the image. The expression is as follows: ； S4.
3. Constructing an optical fuzziness consistency loss Network predicts images An auxiliary pseudo-supervisory image is generated by using an imaging blur degradation model based on the Airy Disk point spread function and adding regional noise through a non-uniform degradation image generation method for industrial scene images. The expression is: ； S4.
4. Construct a composite loss function that includes basic reconstruction loss, structural edge perception loss, and optical blur consistency loss. The resulting expression is: ； in, The weight parameters are those for the structure edge-aware loss. The weighting parameters for optical blur consistency loss are: , ; S5. Acquire industrial scene images, then process them using a multi-view spatial transformation and integration mechanism, and input them into the lightweight neural network trained in step S4 to output an enhanced image of the industrial scene.
2. The industrial image super-resolution reconstruction method based on a physical consistency self-supervised mechanism according to claim 1, characterized in that, The specific implementation method of step S1 includes the following steps: S1.
1. Acquire industrial scene images and perform preprocessing, including channel expansion and median filtering, to obtain preprocessed industrial scene images. ; S1.
2. Then, the preprocessed industrial scene image obtained in step S1.1 is divided into... Given a fixed-size, non-overlapping sub-region, the i-th sub-region is obtained as follows: ; S1.
3. Calculate the average gray value of each sub-region obtained in step S1.2, and obtain the average gray value of the i-th sub-region. ; S1.
4. Based on the average gray value of the sub-region obtained in step S1.3, the Poisson noise intensity coefficient and Gaussian noise intensity coefficient of each region are adaptively adjusted to obtain the expression for the noise generation process: ； in, Generate an industrial scene image for the noise in the i-th sub-region. Represents the Poisson process. Indicates Gaussian noise. Let be the Gaussian noise intensity coefficient of the i-th sub-region. Let be the Poisson noise intensity coefficient of the i-th sub-region. , ; Then, a non-uniform degradation image of the industrial scene is obtained. Represented as: 。 3. The industrial image super-resolution reconstruction method based on a physical consistency self-supervised mechanism according to claim 2, characterized in that, Step S3 constructs a lightweight neural network consisting of three 3×3 convolutional layers and one pixel reconstruction upsampling module. The connection relationship is as follows: input layer → first 3×3 convolutional layer → first ReLU activation layer → second 3×3 convolutional layer → second ReLU activation layer → third 3×3 convolutional layer → pixel shuffle layer → output layer.
4. The industrial image super-resolution reconstruction method based on a physical consistency self-supervised mechanism according to claim 3, characterized in that, In step S4, during the training of the lightweight neural network, the target image is the preprocessed image. The image generated by performing channel averaging, with the input image being... The reconstruction results of the lightweight neural network output are averaged through channels, and then the network is optimized and trained using a composite loss function. After training, the lightweight neural network obtains the reconstruction result from the self-supervised low-resolution image. The reconstruction result is... ,in This represents a trained lightweight neural network. This represents images of an industrial scene that have been collected.
5. The industrial image super-resolution reconstruction method based on a physical consistency self-supervised mechanism according to claim 4, characterized in that, Industrial scene images acquired in step S5 Through a set of geometric transformation operations Then, the results are fed into a pre-trained lightweight neural network to obtain preliminary reconstruction results, followed by inverse transformation. Restored to the original space, the final output image is obtained. The expression is: 。 6. The industrial image super-resolution reconstruction method based on a physical consistency self-supervised mechanism according to claim 5, characterized in that, In step S5, a set of geometric transformation operations are horizontal flip, vertical flip, 90-degree, 180-degree, and 270-degree rotation.