Super-resolution structured light reconstruction method based on wavelet enhanced CNN-Transformer structure

Abstract

The invention discloses a super-resolution structured light reconstruction method based on a wavelet enhanced CNN-Transformer structure. The super-resolution structured light reconstruction method comprises the following steps of: firstly, carrying out wavelet enhancement on a CNN-Transformer structure; according to the method, image reconstruction is realized based on a WECT network adopting a coding-decoding architecture; the method comprises the following steps: firstly, extracting initial features through shallow convolution, performing two-stage down-sampling by using discrete wavelet transform, and respectively inputting results into a convolutional neural network and a Swin Transform module to perform parallel feature extraction and fusion; secondly, performing two-stage up-sampling through inverse discrete wavelet transform, and performing jump connection with shallow layer features; finally, pixel shuffling up-sampling operation is introduced, and residual jump connection between the input image and the output image is established; according to the method, high-resolution details can be efficiently recovered while the structure reduction precision is ensured, the image definition and contrast are enhanced, and the application value and practicability of the SIM image in scenes such as biological imaging and tissue analysis are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing and computer vision, and particularly relates to a structured light illumination microscopic image super-resolution reconstruction method based on a wavelet enhanced CNN-Transformer structure. BACKGROUND

[0002] Structured illumination microscopy (SIM) is an important imaging technology that breaks through the optical diffraction limit and improves spatial resolution, and has wide application value in the fields of cell biology, subcellular structure observation, medical diagnosis, etc. By introducing a structured light illumination mode and combining post-processing reconstruction, SIM can obtain a super-resolution image under a lower light dose. However, in the actual imaging process, limited by the imaging conditions (such as low laser power and high imaging speed), the SIM image usually faces problems such as low signal-to-noise ratio, resolution degradation and structure blur, which puts higher requirements on the image reconstruction algorithm. How to improve the resolution and suppress noise interference while maintaining the structural information of the image is an important direction of current research.

[0003] Existing image reconstruction methods mainly include interpolation methods, transform domain-based reconstruction methods, and deep learning-based super-resolution methods. Among them: interpolation methods (such as bilinear and cubic spline) are simple to implement, but are prone to blurring in edge regions; transform domain-based methods (such as Fourier transform and wavelet transform) have certain frequency expression capabilities, but it is difficult to handle complex textures and multi-scale structures at the same time; deep learning methods have made significant progress in image reconstruction tasks in recent years, especially convolutional neural networks (CNN) and attention mechanism-based models. However, existing deep learning methods still have the following problems in SIM image super-resolution reconstruction:

[0004] (1) Limited feature extraction capability: traditional CNN relies on local convolution operations and is difficult to capture long-distance structure information, while pure Transformer models have global modeling capabilities but lack precise modeling of local details, resulting in the reconstructed image being prone to texture loss or artifacts;

[0005] (2) Severe loss of down-sampling and up-sampling information: most methods use pooling or step convolution for down-sampling, which fails to effectively preserve the multi-frequency features of the image, thereby affecting the final reconstruction quality. SUMMARY

[0006] In view of the deficiencies of the prior art, the purpose of the present application is to provide a wavelet-enhanced CNN-Transformer structure-based super-resolution structured light reconstruction method, which combines the frequency domain decomposition capability of wavelet transform and the local-global modeling advantage of CNN-Transformer, to efficiently recover high-resolution details while ensuring structured reconstruction accuracy, enhance image clarity and contrast, and improve the application value and practicality of SIM images in biological imaging, tissue analysis and other scenarios.

[0007] The wavelet-enhanced CNN-Transformer structure-based super-resolution structured light reconstruction method provided by the present application adopts an encoding-decoding framework, and the overall network structure includes a shallow feature extraction module, a wavelet down-sampling and up-sampling module, a deep feature modeling module, and an image reconstruction module. In the shallow stage, the input low-resolution SIM image is first extracted by the convolution module to extract the initial features; in order to preserve the multi-frequency detail information, the present application introduces discrete wavelet transform (DWT) to replace the traditional pooling operation for two-stage down-sampling. In order to further enhance the feature extraction capability, an integrated convolution-transformer module (ICTB) is proposed, which parallelly integrates a double-scale convolution structure and a Swin Transformer structure, combines the local perception ability of CNN and the global modeling ability of Transformer, while maintaining the compactness of the structure. In the reconstruction stage, inverse discrete wavelet transform (IDWT) is used for feature up-sampling, and the jump connection mechanism between shallow and deep features is used to promote cross-scale information interaction. In order to improve the spatial restoration resolution and image quality, a pixel shuffling up-sampling operation is introduced in the image reconstruction module, and the high-frequency detail recovery capability is enhanced through residual connection. In addition, the present application designs a joint loss function, which combines pixel reconstruction error and structural similarity evaluation index (SSIM) to guide the network optimization process, and takes into account the overall restoration quality and structural detail performance of the image. The technical solutions of the present application are described as follows.

[0008] A wavelet-enhanced CNN-Transformer structure-based super-resolution structured light reconstruction method, comprising the following steps:

[0009] Step 1: Obtain structured light illumination microscopic imaging image data, including a dataset of low-resolution and high-resolution image pairs, and perform preprocessing;

[0010] Step 2: Construct and train a wavelet-enhanced CNN-Transformer super-resolution reconstruction network WECT;

[0011] The wavelet enhancement based CNN-Transformer super-resolution reconstruction network WECT adopts an encoding-decoding architecture and comprises a shallow feature extraction module, a wavelet down-sampling module, a deep feature extraction module, a wavelet up-sampling module and an image reconstruction module.

[0012] The shallow feature extraction module extracts initial features of an input original low-resolution image through shallow convolution.

[0013] The wavelet down-sampling module performs two-stage down-sampling on the extracted initial features using a discrete wavelet transform DWT.

[0014] The deep feature extraction module comprises a double convolution branch in parallel structure for extracting local feature information and a Swin Transformer branch for extracting global feature information, so as to realize parallel feature extraction and fusion of the two-stage down-sampling results of the wavelet down-sampling module and output deep feature maps.

[0015] The wavelet up-sampling module performs two-stage up-sampling on the deep feature maps through an inverse discrete wavelet transform IDWT, restores the spatial size to the original resolution, and is connected to the corresponding wavelet down-sampling module in a skip connection manner to fuse the up-sampling output features and the down-sampling output shallow features.

[0016] The image reconstruction module is used for performing a spatial up-sampling operation on the fused feature maps through a pixel shuffling module, establishing a residual skip connection between the input original low-resolution image and the pixel shuffling module output image, and reconstructing a high-resolution image.

[0017] Step 3: using the wavelet enhancement based CNN-Transformer super-resolution reconstruction network WECT model trained to predict test images and generate high-resolution images.

[0018] In the present application, in step 1, image preprocessing includes the operations of normalization, data enhancement and division of training and test data sets in sequence.

[0019] In the present application, in step 2, the sampling processes of the wavelet down-sampling module and the wavelet up-sampling module both use Haar wavelet bases, wherein:

[0020] The discrete wavelet transform DWT operation decomposes the feature maps into low-frequency and multiple high-frequency components and concatenates them along the channel, which is used to capture texture and structure information.

[0021] The inverse discrete wavelet transform IDWT operation is used to restore the spatial structure of the image, restore the fused feature maps to the original size, and enhance the high-frequency detail restoration capability.

[0022] In the present application, in step 2, the deep feature extraction module is an integrated convolution-transformer module, wherein: a double convolution branch is used to input the input features into 3*3 and 5*5 convolution modules respectively, to extract multi-scale local features, and to enhance the detail retention capability using residual connection; a Swin Transformer branch is used to stack window multi-head self-attention module W-MSA and shift window multi-head self-attention module SW-MSA in turn, to capture cross-region global dependency, and to combine layer normalization LN and feedforward neural network FFN for feature transformation; finally, the output features of the two branches are spliced in the channel dimension, then passed through 1*1 convolution fusion, and then connected with the initial input features to form the output of the deep feature extraction module.

[0023] In the present application, in step 2, in the Swin Transformer branch, window multi-head self-attention module W-MSA and shift window multi-head self-attention module SW-MSA structures are alternately used every two Transformer blocks, respectively used for modeling local self-attention within the window and global interaction information across the window.

[0024] In the present application, in step 2, the pixel shuffling module rearranges the fused feature map into a higher resolution image structure in the channel dimension, and the magnification corresponds to the resolution ratio of the input original low-resolution image and the target image.

[0025] In the present application, in step 2, in the image reconstruction module, the image upsampled by the pixel shuffling module is added to the result of the bilinear interpolation processing of the input original low-resolution image to form a residual jump connection, and a high-resolution image is reconstructed.

[0026] In the present application, in step 2, when training the network, the loss function is composed of pixel reconstruction loss L pixel and structural similarity loss SSIM(Y^,Y) weighted combination, and its expression is:

[0027]

[0028] wherein: L pixel is the mean square error, and λ is the balance coefficient.

[0029] Compared with the prior art, the present application has the following advantages:

[0030] 1) The wavelet transform module is introduced to replace the traditional down-sampling and up-sampling operation, which can more effectively retain the multi-scale structure information of the image and improve the stability and detail performance of the spatial resolution recovery;

[0031] 2) The integrated convolution-transformer module with a parallel structure combines the local feature extraction capability of CNN and the global modeling capability of Transformer, which helps to enhance the reconstruction effect of complex structures and edge details in SIM images;

[0032] 3) The overall network architecture is compact, and the pixel shuffling and residual connection strategy are introduced in the reconstruction module, which improves the reconstruction accuracy while avoiding additional computational overhead, and can realize fast and high-quality SIM image super-resolution reconstruction.

[0033] The present application proposes a wavelet-enhanced CNN-Transformer structure for structural light illumination microscopic image super-resolution reconstruction, which combines the frequency domain decomposition capability of wavelet transform and the local and global modeling advantages of CNN-Transformer. The method can improve the image structure expression capability, effectively restore the detail information, and enhance the adaptability of the model to complex biological samples, providing a new solution for high-quality SIM image reconstruction. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 The specific flowchart of the present application.

[0035] Figure 2 The overall structure diagram of the WECT network proposed by the present application.

[0036] Figure 3 The structure diagram of the integrated convolution-transformer module in the present application.

[0037] Figure 4 The quantitative performance comparison results of the present application in one embodiment for super-resolution reconstruction task.

[0038] Figure 5 The visualization result diagram of the present application in one embodiment for super-resolution reconstruction task.

[0039] Figure 6 The visualization comparison result diagram of the present application in one embodiment and the existing method MIRNet for super-resolution reconstruction under different signal-to-noise ratio conditions. DETAILED DESCRIPTION

[0040] The present application will be described in detail below in conjunction with the accompanying drawings and through specific embodiments. The following described embodiments are an explanation of the present application, and the present application is not limited to the following embodiments.

[0041] The wavelet-enhanced CNN-Transformer structure-based structural light illumination microscopic image super-resolution reconstruction method of the present application has the specific process as shown in Figure 1 The following steps are included:

[0042] Step S1: Obtain structured light illumination microscopic imaging image data, including a dataset of low-resolution and high-resolution image pairs, and pre-process, including normalization, enhancement, and training-test division;

[0043] Step S2: Construct a wavelet-enhanced CNN-Transformer super-resolution reconstruction network WECT, which adopts an encoding-decoding architecture, including a shallow feature extraction module, a wavelet down-sampling and up-sampling module, a deep feature extraction module, and an image reconstruction module;

[0044] Step S3: Model the low-resolution image using the WECT network, first extract initial features through shallow convolution, then use discrete wavelet transform (DWT) for two-stage down-sampling, and input the results to the convolutional neural network (CNN) and Swin Transformer module for parallel feature extraction and fusion; wherein the sampling process uses Haar wavelet basis, and the discrete wavelet transform DWT operation decomposes the feature map into low-frequency and multiple high-frequency components and concatenates along the channel, which is used to capture texture and structure information;

[0045] Step S4: Restore the spatial size by two-stage up-sampling through inverse discrete wavelet transform (IDWT), and perform jump connection with shallow features; wherein the sampling process uses Haar wavelet basis, and the IDWT operation is used to restore the image spatial structure, restore the fused feature map to the original size, and enhance the high-frequency detail restoration capability;

[0046] Step S5: Introduce pixel shuffling up-sampling operation in the image reconstruction module, and establish residual jump connection between the input image and the output image, which is used to further improve the reconstruction accuracy and texture performance;

[0047] Step S6: Set the training parameters and loss function of the network, combine the pixel reconstruction loss and structural similarity loss for optimization training, and finally complete the reconstruction process of high-quality SIM images; The parameter configuration in the network training process includes model parameters and training parameters; The model parameters include the stacking depth, channel number setting and attention head number of the deep feature extraction module; The training parameters include batch size, initial learning rate and training period;

[0048] Step S7: Using the trained WECT model to predict the test image, generate a high-resolution image, and evaluate and visually analyze the performance of the reconstructed image through multiple indicators such as peak signal-to-noise ratio (PSNR), structural similarity (SSIM), multi-scale structural similarity (MS-SSIM), and normalized mean square error (NRMSE). Among them, the peak signal-to-noise ratio (PSNR) is used to evaluate the pixel fidelity of the image; the structural similarity (SSIM) and the multi-scale structural similarity (MS-SSIM) are used to evaluate the consistency of the image structure and the detail restoration ability; the normalized mean square error (NRMSE) is used to measure the pixel difference between the prediction result and the real image.

[0049] Specifically, in step S1, the structured light illumination microscopic image dataset is obtained and preprocessed, including the following steps:

[0050] (1) Select the structured light illumination super-resolution image dataset BioSR, which covers four typical biological structures: CCP, ER, MT, and F-actin. All 2D image pairs are cropped to image blocks with sizes of 128x128 and 256x256, respectively, as input and output samples of the model.

[0051] (2) The percentile normalization method is used to map the pixel values of the image samples to the range of [0, 1], and the calculation formula is as follows:

[0052]

[0053] Where perc(u, p) represents the p-th percentile of all pixel values of the input image u.

[0054] (3) Data augmentation is performed on the normalized images, including random horizontal flipping, random vertical flipping, and angle rotation, to improve the generalization ability and robustness of the model;

[0055] (4) The processed image blocks are divided into training set and test set according to the ratio of 8:2, ensuring that the low-resolution image and its corresponding high-resolution image maintain consistent pairing.

[0056] In step S2, an integrated convolution-transformer module (ICTB) for deep feature extraction is constructed, and its structure is as follows: Figure 3The ICTB module is composed of a double convolution branch and a Swin Transformer branch in parallel, which are used to extract local and global feature information, respectively. First, the input feature map from the DWT module is set to 128 channels, and is evenly divided into two sub-branches along the channel direction, with 64 channels for each sub-branch. The double convolution branch includes two parallel convolution modules using 3x3 and 5x5 convolution kernels, respectively. Each convolution module contains two consecutive convolution layers with 64 convolution channels, followed by a ReLU activation function. The double convolution branch extracts multi-scale local features and uses residual connection to enhance detail preservation. In the Swin Transformer branch, window multi-head self-attention modules (W-MSA) and shift window multi-head self-attention modules (SW-MSA) are stacked in sequence to capture cross-region global dependencies and combine layer normalization (LN) and feed-forward neural networks (FFN) for feature transformation. W-MSA and SW-MSA structures are alternately used in every two Transformer blocks to model local self-attention within the window and global interaction information across windows, respectively. Each Transformer block contains a feed-forward neural network (FFN), a layer normalization (LN), and a residual connection module, and its feature update process is as follows:

[0057]

[0058] Specifically, the Swin Transformer branch receives another part of the 64-channel feature map and inputs it into two Swin Transformer blocks. Each block contains a window size of 7x7 multi-head self-attention module (W-MSA) and a shift window multi-head self-attention module (SW-MSA), with 4 attention heads and a feature embedding dimension of 64. In each Transformer module, a feed-forward neural network (FFN) module is configured with a hidden layer channel number of 128, and residual connection and layer normalization are used to improve model stability.

[0059] Finally, the features extracted by the double convolution path and the Transformer path are concatenated and fused through a 1x1 convolution, outputting 128 channels, and forming a residual connection with the original input feature map.

[0060] In step S3, the low-resolution image is modeled using the WECT network to extract multi-scale deep features. First, the input image is extracted through a 3x3 convolution layer to extract shallow features, with 64 channels and a ReLU activation function. Next, two-level discrete wavelet transform (DWT) is used for downsampling, reducing the spatial size of the feature map to 1 / 2 and 1 / 4 of the original, respectively, while preserving the low-frequency and high-frequency information through channel concatenation. Subsequently, the feature map is input into the integrated convolution-transformer module (ICTB) with a parallel structure, which models local and global features through a double-scale convolution path and a Swin Transformer path, respectively, and outputs the fused deep feature representation. The shallow features are preserved through a skip connection and fused with the deep features in the subsequent reconstruction stage to enhance the detail fidelity.

[0061] In step S4, the deep feature map is first input into two IDWT modules in sequence for upsampling, gradually restoring the spatial size to the original resolution. Each level of IDWT uses the Haar wavelet basis and combines the structure and details of the low-frequency and high-frequency sub-band reconstructed feature maps. Subsequently, the upsampled feature map and the preserved shallow features from the S3 stage are fused through a skip connection and then passed through a 3x3 convolution layer (64 channels) to unify the feature channel dimension, providing complete feature support for the final image reconstruction.

[0062] In step S5, the fused feature map is input into the Pixel Shuffle module for the final spatial upsampling to reconstruct the high-resolution image. The Pixel Shuffle module rearranges the channel information in the feature map into a higher-resolution image structure with a magnification of 2x, corresponding to the resolution ratio of the input low-resolution image and the target image. To improve the fidelity and texture performance of the reconstructed image, the Pixel Shuffle output image is added to the bilinear upsampling result of the input image to construct a residual connection:

[0063]

[0064] wherein, represents the final reconstructed image, F out is the feature map output by the Pixel Shuffle module, and X is the original input image.

[0065] In step S6, the AdamW optimizer is used in the training stage, with an initial learning rate of 4e-4, a batch size of 16, and 50 epochs (Epoch) for training. The cosine annealing scheduler is used to gradually reduce the learning rate to improve the convergence stability of the model. The training platform configuration includes: dual NVIDIA RTX 3090 GPUs (24GB video memory), Intel i9-10980XE CPU, 128GB memory. The loss function is composed of the pixel reconstruction loss and the structural similarity loss, defined as:

[0066]

[0067] wherein L pixel is the mean square error, SSIM(Y^,Y) is the structural similarity index, used to enhance the structural information perception, and λ is the balance coefficient, set to 0.1. The combined loss ensures pixel accuracy while improving the structural fidelity and detail performance of the reconstructed image.

[0068]

[0069] wherein, Y respectively represents the predicted image obtained using the training model and the real image, μ Y respectively represents the mean of the image and the image Y, σ Y respectively represents the standard deviation of the image and the image Y, represents the covariance of the image and the image Y, and C1 and C2 are constants.

[0070] In step S7, the WECT model trained is used to perform super-resolution reconstruction on the low-resolution images in the test set, and the performance of the reconstructed images is quantitatively and qualitatively analyzed. In this embodiment, the network parameters are saved once every 500 iterations during the training process, which is used for subsequent model selection and performance evaluation. Finally, the model with the optimal validation index is used to infer the test set to generate high-resolution images.

[0071] As Figure 4 shown, the quantitative performance comparison results of the method of the present application and six representative deep learning models on the BioSR dataset. The evaluation indicators include peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), multi-scale structural similarity (MS-SSIM), and normalized root mean square error (NRMSE), which comprehensively measure the performance of the model in terms of reconstruction quality, structure restoration, and noise suppression. It can be seen that the method of the present application has achieved leading results in each indicator, especially in the SSIM and MS-SSIM indicators, showing better structural fidelity capability, indicating that the proposed WECT network has stronger high-resolution reconstruction capability and microscopic detail restoration effect.

[0072] In addition, qualitative visual analysis is performed, Figure 5 which shows the reconstruction effect of the method of the present application on the "endoplasmic reticulum (ER)" structure image in the BioSR dataset. As can be seen from the figure, the super-resolution image generated by WECT is significantly better than the original low-resolution input in terms of structure detail restoration and edge clarity, reflecting good texture fidelity.

[0073] Further, as shown in Figure 6 Table 6, the super-resolution reconstruction effects of MIRNet and the WECT method of the present application under three different signal-to-noise ratio (SNR) conditions are compared. The results show that WECT still has stronger detail recovery capability and visual consistency under low SNR conditions, indicating that the method has stronger robustness and generalization performance in complex imaging environments and can meet the high-quality biological imaging super-resolution reconstruction requirements.

Claims

1. A wavelet-enhanced CNN-Transformer structure-based super-resolution structured light reconstruction method, characterized in that, The method comprises the following steps: Step 1: obtaining structured light illumination microscopic imaging image data, including a data set of a low-resolution and high-resolution image pair, and preprocessing; Step 2: constructing and training a wavelet-enhanced CNN-Transformer super-resolution reconstruction network WECT; The wavelet-enhanced CNN-Transformer super-resolution reconstruction network WECT adopts an encoding-decoding architecture, and comprises a shallow feature extraction module, a wavelet down-sampling module, a deep feature extraction module, a wavelet up-sampling module and an image reconstruction module; wherein: The shallow feature extraction module extracts initial features of the input original low-resolution image through shallow convolution; The wavelet down-sampling module uses discrete wavelet transform (DWT) to perform two-stage down-sampling on the extracted initial features; The deep feature extraction module comprises a double convolution branch for extracting local feature information and a Swin Transformer branch for extracting global feature information in a parallel structure, so as to realize parallel feature extraction and fusion of the two-stage down-sampling results of the wavelet down-sampling module, and output deep feature maps; The wavelet up-sampling module performs two-stage up-sampling on the deep feature maps through inverse discrete wavelet transform (IDWT), restores the spatial size to the original resolution, and is connected to the corresponding wavelet down-sampling module to fuse the up-sampling output features and the down-sampling output shallow features; The image reconstruction module is used for spatial up-sampling operation of the fused feature maps through a pixel shuffling module, and establishes a residual skip connection between the input original low-resolution image and the pixel shuffling module output image, to reconstruct a high-resolution image; Step 3: using the trained wavelet-enhanced CNN-Transformer super-resolution reconstruction network WECT model to predict test images to generate high-resolution images.

2. The super-resolution structured light reconstruction method of claim 1, wherein, In step 1, the image preprocessing includes the operations of normalization, data enhancement and division of training and test data sets in sequence.

3. The super-resolution structured light reconstruction method of claim 1, wherein, In step 2, the sampling processes of the wavelet down-sampling module and the wavelet up-sampling module both use Haar wavelet bases, wherein: The discrete wavelet transform (DWT) operation decomposes the feature maps into low-frequency and multiple high-frequency components and concatenates them along the channel, for capturing texture and structure information; The inverse discrete wavelet transform (IDWT) operation is used for restoring the spatial structure of the image, restoring the fused feature maps to the original size, and enhancing the high-frequency detail restoration capability.

4. The super-resolution structured light reconstruction method of claim 1, wherein, In step 2, the deep feature extraction module is an integrated convolution-transformer module, wherein: a double convolution branch is used to input the input features into 3*3 and 5*5 convolution modules respectively, to extract multi-scale local features, and to enhance the detail preservation ability using residual connection; a SwinTransformer branch is used to stack window multi-head self-attention module W-MSA and shift window multi-head self-attention module SW-MSA in turn, to capture cross-region global dependency, and to combine layer normalization LN and feedforward neural network FFN for feature transformation; finally, the output features of the two branches are concatenated in the channel dimension, then passed through 1*1 convolution fusion, and then connected with the initial input features to form the output of the deep feature extraction module.

5. The super-resolution structured light reconstruction method of claim 1, wherein, In step 2, in the SwinTransformer branch, window multi-head self-attention module W-MSA and shift window multi-head self-attention module SW-MSA structures are alternately used between every two Transformer blocks, respectively used for modeling local self-attention within the window and global interaction information across the window.

6. The super-resolution structured light reconstruction method of claim 1, wherein, In step 2, the pixel shuffling module rearranges the fused feature map in the channel dimension into a higher resolution image structure, and the magnification corresponds to the resolution ratio of the input original low-resolution image and the target image.

7. The super-resolution structured light reconstruction method of claim 1, wherein, In step 2, in the image reconstruction module, the image upsampled by the pixel shuffling module is added to the result of the bilinear interpolation processing of the input original low-resolution image, to form a residual jump connection, and to reconstruct a high-resolution image.

8. The super-resolution structured light reconstruction method of claim 1, wherein, In step 2, when training the network, the loss function is composed of pixel reconstruction loss L pixel and structural similarity loss SSIM(Y^,Y) weighted combination, the expression is: where: L pixel is the mean square error, and λ is the balancing coefficient.

9. The super-resolution structured light reconstruction method of claim 1, wherein, In step 2, the parameter configuration in the network training process includes model parameters and training parameters; the model parameters include the stacking depth, channel number setting and attention head number of the deep feature extraction module; The training parameters include batch size, initial learning rate and training period.

10. The super-resolution structured light reconstruction method of claim 1, wherein, In step 3, the performance of the reconstructed image is further evaluated and visually analyzed through peak signal-to-noise ratio PSNR, structural similarity SSIM, multi-scale structural similarity MS-SSIM and normalized root mean square error NRMSE.

Images