Industrial battery CT image super-resolution method based on deep learning

Abstract

The application belongs to the field of CT tomographic reconstruction and the field of artificial intelligence, and discloses an industrial battery CT image super-resolution method based on deep learning. The method takes a low-resolution tomographic image as input, constructs an end-to-end battery CT image super-resolution network, and outputs a high-resolution tomographic image. The network uses multi-scale attention module (MSA) and enhanced grouping module (EGM) for parallel calculation, adopts hybrid channel attention feed-forward module (HCFB) for nonlinear enhancement, and combines a frequency modulation module for frequency domain feature modeling. The method can obtain a battery CT tomographic reconstruction result with clearer boundaries, more coherent layer structures and fewer artifacts, and provides a more stable and reliable data basis for subsequent defect detection, structure segmentation and three-dimensional analysis.

Description

Technical Field

[0001] This invention belongs to the fields of industrial CT and artificial intelligence, and relates to a deep learning-based super-resolution method for industrial battery CT images. Background Technology

[0002] CT technology, capable of reconstructing the intricate three-dimensional structure of objects, has been widely used for detecting internal defects and assessing structural quality in industrial products, including complex components, welds, and castings. For objects with fine structures and small-scale defects, such as power batteries, smaller voxel sizes, higher geometric magnification, and more adequate projection sampling and exposure conditions are typically required to obtain high-resolution tomographic images to identify defects like electrode interfaces, microcracks, and pores. This often falls short of time constraints, leading to offline sampling being the primary method for battery CT inspection in many scenarios. To balance efficiency and quality, practical applications often employ methods such as reducing resolution and shortening exposure times for CT scan reconstruction. However, the resulting loss of detail, blurred boundaries, and noise artifacts directly impact the accuracy and stability of subsequent quality inspections. Therefore, improving the resolution of tomographic images without significantly increasing scan time has become a key challenge for the practical application of industrial battery CT.

[0003] Industrial battery CT images simultaneously contain local micro-defects in the electrodes and layered boundary structures. The layered boundaries are highly sensitive to global geometric consistency. Therefore, battery CT super-resolution not only requires the model to have the ability to restore local details, but also requires it to maintain the continuity and consistency of the overall layered structure during large-scale information fusion. At the same time, CT images are often accompanied by noise and stripe artifacts, which can easily lead to boundary breaks, texture misalignment, and false textures, thereby reducing the interpretability and engineering usability of the results. Existing work often employs deep learning super-resolution models in the image domain, achieving resolution enhancement and detail restoration by learning the mapping relationship from low-resolution to high-resolution images. However, CNN-based approaches are limited by local receptive fields, often struggling to establish stable global constraints for large-scale extended layered structures, and are prone to structural continuity degradation at interlayer boundaries. While Transformer-based approaches possess stronger global modeling capabilities, in the presence of strong noise and stripe artifacts, self-attention struggles to simultaneously address detail restoration and artifact suppression, leading to excessive texture enhancement or structural distortion. To model long-range dependencies between spatial locations in images, state-space model-based approaches recursively update hidden states along the scanning direction, thereby enabling information transfer and correlation between distant locations. However, in scenarios like battery CT with layered structures and significant degenerative coupling, the problem of insufficient local detail restoration and inadequate global geometric consistency constraints may still exist, making it difficult to simultaneously meet the engineering requirements of defect discernibility and structural reliability.

[0004] Therefore, exploring a super-resolution method for battery CT images that can restore details in CT tomographic images while maintaining the global consistency of the layered structure and effectively suppress noise and artifacts from interfering with the boundaries is of significant research value. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing technologies and provide a super-resolution reconstruction method for CT tomographic images in the power battery industry. This method uses a multi-scale attention module (MSA) and an enhancement grouping module (EGM) for parallel computation, employs a hybrid channel attention feedforward module (HCFB) for nonlinear enhancement, and combines a frequency modulation module for frequency domain feature modeling, thereby improving the detail recovery quality and structural stability of battery CT tomographic images.

[0006] The technical solution of the present invention:

[0007] A deep learning-based super-resolution method for industrial battery CT images includes the following steps:

[0008] Step 1: Construct the battery CT dataset;

[0009] Multiple battery samples of the same model were prepared, and each sample was scanned using the same CT scanner to obtain projection data. During the scanning process, the clamping method, spatial orientation, and starting angle of the battery samples were kept consistent. The scanning angle step was 0.1 degrees, and the acquisition frame rate was 1 frame. The acquired projection data was reconstructed using the FDK reconstruction algorithm to obtain CT tomographic images with a resolution of 396×972. The high-resolution ground truth (HR) data is used as the basis for the analysis. Subsequently, the CT tomographic images are downsampled at a preset magnification and Gaussian noise is added to generate the corresponding low-resolution CT tomographic images. ;

[0010] The battery CT dataset includes CT tomographic images with a resolution of 396×972. Corresponding low-resolution CT tomographic images Supervised training and validation of super-resolution networks for battery CT images;

[0011] Step 2: Construct a super-resolution network for battery CT images;

[0012] The super-resolution network for industrial battery CT images consists of a shallow feature extraction module, a deep feature extraction module, and an upsampling module.

[0013] The shallow feature extraction module uses low-resolution CT tomographic images As input, its number of channels The number is 3; the shallow feature extraction module uses a 3×3 convolutional layer to embed features into the input, mapping the CT tomographic image to the feature space to obtain the number of channels. Shallow features of 180 ;

[0014] The deep feature extraction module receives the shallow features output by the shallow feature extraction module. Furthermore, within the deep feature extraction module, each residual attention state space module maintains a constant number of input and output channels. The deep feature extraction module consists of The residual attention state space modules are sequentially stacked in series, and the stacked module is composed of the first... The output of each residual attention state space module is denoted as the deep feature. ,in After stacking, the final deep features are obtained. And deep features are connected through long skip connections. shallow features Features are obtained by adding and fusing elements one by one. As the output of the deep feature extraction module, it is used for reconstruction by the subsequent upsampling module;

[0015] Among them, the +1 residual attention state space module with the first Deep features output by each residual attention state space module As input, and within the residual attention state space module, the number of input channels and the number of output channels are kept constant. The residual attention state space module consists of The attention state space modules are sequentially stacked in series, the first... The output of each attention state space module is a feature ,in After stacking, the output features are obtained. Then output features The input frequency modulation module (FMB) obtains frequency domain modulation features, which are then followed by a 3×3 convolutional layer for feature integration. Finally, the output of the 3×3 convolutional layer is combined with the deep features via residual connections. Adding each element one by one, we get the first... Output features of +1 residual attention state space module .

[0016] The frequency modulation module specifically includes: first, the input features are transformed to the frequency domain by FFT, and the real and imaginary parts are concatenated in the channel dimension and then modeled in the frequency domain by Dwconv, and residual connections are used to preserve the spectral information; then, the frequency domain features are respectively fed into the high-frequency branch and the low-frequency branch for branch processing. The high-frequency branch extracts and modulates high-frequency related components by 3×1 convolution and 1×3 convolution in sequence, and the low-frequency branch extracts and modulates low-frequency related components by 3×3 convolution; the outputs of the high-frequency branch and the low-frequency branch are fused and fed into Pwconv and SiLU for integration, and finally back-mapped to the spatial domain by IFFT to obtain the output of the frequency modulation module.

[0017] The first +1 attention state space module with the first Features output by each attention state space module As input, first the features Layer-normalized LN processing is performed; the features after layer-normalized LN processing are then input into the multi-scale attention module (MSA) and the enhancement group Mamba module (EGM) for parallel computation, and the outputs of the MSA and EGM are added together to obtain the fused output. Subsequently, the features are... Multiplying by a scaling factor (scale) yields the scaled feature, which is then added to the fused output to obtain the intermediate feature. The intermediate feature is then subjected to Layer Normalization (LN) processing and input into the Hybrid Channel Attention Feedforward (HCFB) module to obtain the feedforward output. Finally, the intermediate feature is multiplied by the scaling factor (scale) to obtain the scaled feature, which is then added to the feedforward output to obtain the output feature of the attention state space module. .

[0018] The multi-scale attention module specifically includes: inputting the input features sequentially into three parallel multi-head neighborhood attention branches, with the neighborhood window size of each multi-head neighborhood attention branch set to 11×11 and the hole ratios set to 1, 2, and 4 respectively; each multi-head neighborhood attention branch performing multi-head neighborhood attention calculation on the input features to obtain the output features; then fusing the output features of the three multi-head neighborhood attention branches in the channel dimension, and integrating the channels through a 1×1 convolution layer to obtain the output of the multi-scale attention module.

[0019] The enhanced grouping Mamba module specifically includes: first, inputting the input features into a linear mapping layer for channel transformation; then, performing branch splitting in the channel dimension to form two feature paths, one of which serves as a gated branch, undergoing only SiLU activation to obtain gate weights; the other feature path serves as the main branch, sequentially extracting local spatial features through Dwconv and enhancing the nonlinear expression through SiLU to obtain intermediate features; subsequently, the intermediate features are divided into four groups of features according to channels by the grouping two-dimensional state space G-SS2D, and the pixels of the four groups of features are sequentially scanned and state recursively processed along the four directions to obtain grouped features, which are then processed by the grouping space-channel hybrid module GSCMixer for spatial transformation. Spatial and channel mixing is performed by rearranging grouped features along the channel dimension and dividing them into groups, so that each group of features contains responses from different directions simultaneously. Then, 3×3 group convolution is used to complete the spatial mixing within the group to obtain spatial mixed features. Global average pooling and 1×1 group convolution are then performed on the spatial mixed features to obtain channel features. One-dimensional state space modeling is then applied to the channel features to generate a gating signal and modulate the spatial mixed features. Finally, inter-group information exchange is achieved through channel rearrangement and output. Finally, layer normalization (LN) is used to stabilize the feature distribution. The output of the enhanced grouped Mamba module fuses the results of the gating branch and the main branch through element-wise multiplication to obtain the output features of the enhanced grouped Mamba module.

[0020] The hybrid channel attention feedforward module specifically includes: first, feeding the input features into the MsConv hybrid convolutional unit, with convolutional branches of depthwise convolutional kernel sizes of 1×1, 3×3, and 5×5 being computed in parallel, and concatenating the outputs of each depthwise convolutional branch to obtain fused features; then, performing two branches on the fused features in the channel dimension, one branch being activated by GELU to obtain a gated branch, and the other branch serving as the main branch, with the results of the two branches being multiplied element-wise; finally, channel integration is performed through a PWConv layer, and channel attention CA is connected to obtain the output of the hybrid channel attention feedforward module.

[0021] The upsampling module receives features output from the deep feature extraction module. First, regarding the features Perform a 3×3 convolution to complete the channel expansion, and set the magnification to be [value missing]. Then the number of output channels of the 3×3 convolution is set to The 3×3 convolution output is then fed into the PixelShuffle layer, magnified according to the magnification factor. Rearrange the sub-pixels in the channel dimension to the spatial dimension to obtain a spatial resolution magnified to... And the number of channels is The features are then reconstructed using 3×3 convolution, mapping the upsampled features to the number of channels. High-resolution CT images ;

[0022] Step 3: Train the super-resolution network for battery CT images;

[0023] The training of the battery CT image super-resolution network mainly involves training three parts: shallow feature extraction, deep feature extraction, and upsampling. The battery CT dataset obtained in step 1 is divided into a training set and a test set, with a ratio of 7:1. Horizontal and vertical flipping are used for data augmentation during training. During training, a low-resolution CT tomographic image is randomly selected from the training set as input to the battery CT image super-resolution network. After shallow feature extraction, it enters deep feature extraction, and finally, the upsampling module generates a high-resolution CT image at the corresponding magnification. Subsequently, a loss function was used to calculate high-resolution CT images. CT tomographic images with a resolution of 396×972 The loss between the two is calculated, and the parameters are updated using Adam; the loss function consists of the mean absolute error (MAE).

[0024]

[0025] exist After convergence, the performance of the battery CT image super-resolution network was tested using a test set to verify its generalization ability. The parameters of the battery CT image super-resolution network were then saved to the model file for subsequent fast inference.

[0026] Step 4: Execution of the super-resolution network for battery CT images;

[0027] When testing battery samples, the battery sample is first scanned using a CT scanner to complete preliminary tomographic reconstruction and obtain a tomographic image with a complete structure but limited resolution. Then, the tomographic image is sent to the battery CT image super-resolution network built in step 2, and the parameters of the battery CT image super-resolution network model file saved in step 3 are imported to perform inference to obtain the corresponding high-resolution CT image, which is used for subsequent defect detection, structural segmentation and three-dimensional analysis.

[0028] The beneficial effects of this invention are:

[0029] (1) This invention proposes a super-resolution reconstruction method for CT tomographic images of power battery industry, which improves resolution while restoring details and effectively maintains the continuity and consistency of layered structure and quasi-periodic texture over a wide range.

[0030] (2) This invention introduces multi-scale attention and two-dimensional state space model mechanism into the network, taking into account both the detail enhancement of local edges and the transmission constraints of long-distance layered structures, thereby reducing structural distortions such as boundary breakage and texture misalignment and improving reconstruction stability.

[0031] (3) Modeling features in the frequency domain can reduce the interference of noise and fringe artifacts on high-frequency information, reduce the risk of over-enhancement and pseudo-detail generation, thereby improving the stability and engineering usability of the reconstruction results. Attached Figure Description

[0032] Figure 1 This is the overall architecture of the battery CT image super-resolution network of the present invention;

[0033] Figure 2 This is a block diagram of the attention state space of the present invention;

[0034] Figure 3 This is a diagram of the frequency modulation module of the present invention;

[0035] Figure 4 This is a diagram of the grouped space-channel hybrid module of the present invention. Detailed Implementation

[0036] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings and technical solutions.

[0037] First, based on the battery CT dataset established in step 1, 40 battery samples of the same model were acquired. Then, projection data was obtained by scanning each battery sample, and CT tomographic images with a resolution of 396×972 were reconstructed using the FDK reconstruction algorithm. As the truth value; subsequently, CT tomographic images with a resolution of 396×972 were analyzed. By downsampling at a preset magnification and adding Gaussian noise, a low-resolution CT tomographic image corresponding to the true value is generated. Thus, a pairwise battery CT dataset is constructed for supervised learning;

[0038] Secondly, a battery CT image super-resolution network is constructed according to step 2; the overall architecture of the battery CT image super-resolution network is as follows: Figure 1 As shown, it consists of a shallow feature extraction module, a deep feature extraction module, and an upsampling module; the shallow feature extraction module uses a 3×3 convolutional layer to extract CT tomographic images. Mapping to the feature space yields shallow features. The deep feature extraction module receives shallow features. Deep features are obtained through six cascaded residual attention state space modules. Each residual attention state space module consists of 6 attention state space modules and 1 frequency modulation module, as shown in the diagram. Figure 2 As shown, the input features are first subjected to Layer Normalization (LN) processing. The LN-processed features are then input into the Multi-Scale Attention (MSA) module and the Enhancement Grouping (EGM) module for parallel computation. The outputs of the MSA and EGM modules are then added to obtain a fused output. Next, the input features are multiplied by a scaling factor (scale) to obtain scaled features, which are then added to the fused output to obtain intermediate features. These intermediate features are then subjected to LN processing and input into the Hybrid Channel Attention Feedforward (HCFB) module to obtain a feedforward output. Finally, the intermediate features are multiplied by a scaling factor (scale) to obtain scaled features, which are then added to the feedforward output to obtain the output features of the attention state space module. The frequency modulation module is as follows: Figure 3 As shown, the input features are first transformed to the frequency domain by FFT. The real and imaginary parts are concatenated along the channel dimension and then modeled using Dwconv, with residual connections used to preserve spectral information. The frequency domain features are then fed into high-frequency and low-frequency branches for branching processing. The high-frequency branch extracts and modulates high-frequency related components through 3×1 and 1×3 convolutions, while the low-frequency branch extracts and modulates low-frequency related components through 3×3 convolutions. The outputs of the high-frequency and low-frequency branches are fused and fed into Pwconv and SiLU for integration. Finally, the output is back-mapped to the spatial domain by IFFT to obtain the output of the frequency modulation module. The upsampling module receives the deep features. First, regarding the features A 3×3 convolution is performed to expand the channels. The output of the 3×3 convolution is then fed into a PixelShuffle layer. Finally, a 3×3 convolution is used for reconstruction mapping, mapping the upsampled features to the target magnification to output a high-resolution CT image. ;

[0039] Then, the battery CT image super-resolution network is trained according to step 3. First, the paired battery CT dataset of the above 40 battery samples is divided into training and testing sets in a 7:1 ratio, i.e., 35 samples are used for training and 5 samples are used for testing, to verify the training effect of the battery CT image super-resolution network. The battery CT image super-resolution network is trained using an RTX 4090 GPU, and the change of its loss function Loss is observed. The battery CT image super-resolution network training is completed when the Loss stabilizes and converges in about 12 hours. At this time, the battery CT image super-resolution network parameters are saved to the model file, which can be imported when inference is needed later.

[0040] After completing the above steps, you can proceed with the actual use according to step 4. When it is necessary to test the battery sample, first use a CT device to scan and reconstruct a low-resolution CT tomographic image, and then send it into the battery CT image super-resolution network. Call the network parameters saved in the model file to perform inference and generate a clear high-resolution CT tomographic image.

Claims

1. A deep learning-based super-resolution method for industrial battery CT images, characterized in that, Includes the following steps: Step 1: Construct the battery CT dataset; Multiple battery samples of the same model were prepared, and each sample was scanned using the same CT scanner to obtain projection data. During the scanning process, the clamping method, spatial orientation, and starting angle of the battery samples were kept consistent. The scanning angle step was 0.1 degrees, and the acquisition frame rate was 1 frame. The acquired projection data was reconstructed using the FDK reconstruction algorithm to obtain CT tomographic images with a resolution of 396×972. The high-resolution ground truth (HR) data is used as the basis for the analysis. Subsequently, the CT tomographic images are downsampled at a preset magnification and Gaussian noise is added to generate the corresponding low-resolution CT tomographic images. ; The battery CT dataset includes CT tomographic images with a resolution of 396×972. Corresponding low-resolution CT tomographic images Supervised training and validation of super-resolution networks for battery CT images; Step 2: Construct a super-resolution network for battery CT images; The super-resolution network for industrial battery CT images consists of a shallow feature extraction module, a deep feature extraction module, and an upsampling module. The shallow feature extraction module uses low-resolution CT tomographic images As input, its number of channels It is 3; The shallow feature extraction module uses a 3×3 convolutional layer to embed features into the input, mapping the CT tomographic image to the feature space to obtain the number of channels. Shallow features of 180 ; The deep feature extraction module receives the shallow features output by the shallow feature extraction module. Furthermore, within the deep feature extraction module, each residual attention state space module maintains a constant number of input and output channels. ; The deep feature extraction module consists of The residual attention state space modules are sequentially stacked in series, and the stacked module is composed of the first... The output of each residual attention state space module is denoted as the deep feature. ,in After stacking, the final deep features are obtained. And deep features are connected through long skip connections. shallow features Features are obtained by adding and fusing elements one by one. As the output of the deep feature extraction module, it is used for reconstruction by the subsequent upsampling module; No. +1 residual attention state space module with the first Deep features output by each residual attention state space module As input, and within the residual attention state space module, the number of input channels and the number of output channels are kept constant. The residual attention state space module consists of The attention state space modules are sequentially stacked in series, the first... The output of each attention state space module is a feature ,in After stacking, the output features are obtained. Then output features The input frequency modulation module (FMB) obtains frequency domain modulation features, which are then followed by a 3×3 convolutional layer for feature integration. Finally, the output of the 3×3 convolutional layer is combined with the deep features via residual connections. Adding each element one by one, we get the first... Output features of +1 residual attention state space module ; The upsampling module receives features output from the deep feature extraction module. First, regarding the features Perform a 3×3 convolution to complete the channel expansion, and set the magnification to be [value missing]. Then the number of output channels of the 3×3 convolution is set to ; The 3×3 convolution output is then fed into the PixelShuffle layer, magnified according to the magnification factor. Rearrange the sub-pixels in the channel dimension to the spatial dimension to obtain a spatial resolution magnified to... And the number of channels is The features are then reconstructed using 3×3 convolution, mapping the upsampled features to the number of channels. High-resolution CT images ; Step 3: Train the super-resolution network for battery CT images; Training a super-resolution network for battery CT images includes training shallow feature extraction, deep feature extraction, and upsampling modules; The battery CT dataset obtained in step 1 is divided into a training set and a test set, with the ratio of the training set to the test set set set to 7:1; at the same time, horizontal flipping and vertical flipping are used to augment the data during the training process. During training, a low-resolution CT tomographic image is randomly selected from the training set as input to the battery CT image super-resolution network. After shallow feature extraction, it enters deep feature extraction, and finally, the upsampling module generates a high-resolution CT image at the corresponding magnification. Subsequently, a loss function was used to calculate high-resolution CT images. CT tomographic images with a resolution of 396×972 The loss between the two is calculated, and the parameters are updated using Adam; the loss function consists of the mean absolute error (MAE). exist After convergence, the performance of the battery CT image super-resolution network was tested using a test set to verify its generalization ability. The parameters of the battery CT image super-resolution network were then saved to the model file for subsequent fast inference.

2. The deep learning-based super-resolution method for industrial battery CT images according to claim 1, characterized in that, No. +1 attention state space module with the first Features output by each attention state space module As input, first the features Layer-normalized LN processing is performed; the features after layer-normalized LN processing are then input into the multi-scale attention module (MSA) and the enhancement group Mamba module (EGM) for parallel computation, and the outputs of the MSA and EGM are added together to obtain the fused output. Subsequently, the features are... Multiplying by a scaling factor (scale) yields the scaled feature, which is then added to the fused output to obtain the intermediate feature. The intermediate feature is then subjected to Layer Normalization (LN) and input into the Hybrid Channel Attention Feedforward (HCFB) module to obtain the feedforward output. Finally, the intermediate feature is multiplied by the scaling factor (scale) to obtain the scaled feature, which is then added to the feedforward output to obtain the output feature of the attention state space module. .

3. The deep learning-based super-resolution method for industrial battery CT images according to claim 2, characterized in that, The multi-scale attention module specifically includes: The input features are sequentially fed into three parallel multi-head neighborhood attention branches. The neighborhood window size of each multi-head neighborhood attention branch is set to 11×11, and the hole ratios are set to 1, 2, and 4, respectively. Each multi-head neighborhood attention branch performs multi-head neighborhood attention calculation on the input features to obtain the output features. Then, the output features of the three multi-head neighborhood attention branches are fused in the channel dimension, and channel integration is performed through a 1×1 convolution layer to obtain the output of the multi-scale attention module.

4. The deep learning-based super-resolution method for industrial battery CT images according to claim 2, characterized in that, The enhanced group Mamba module specifically includes: First, the input features are fed into a linear mapping layer for channel transformation. Then, branching is performed along the channel dimension to form two feature paths. One path serves as a gated branch, undergoing only SiLU activation to obtain gate weights. The other path serves as the main branch, sequentially extracting local spatial features via Dwconv and enhancing the nonlinear expression with SiLU to obtain intermediate features. Subsequently, the intermediate features are divided into four groups according to channels using the grouped 2D state space G-SS2D. Sequential scanning and state recursion are then performed on the pixels of each of the four groups to obtain grouped features. Finally, the grouped spatial-channel mixing module GSCMixer performs spatial and channel mixing, further refining the grouped features. Features are rearranged and grouped along the channel dimension, so that each group of features contains responses from different directions. Then, 3×3 grouped convolution is used to perform spatial mixing within the group to obtain spatial mixed features. Global average pooling and 1×1 grouped convolution are then performed on the spatial mixed features to obtain channel features. One-dimensional state space modeling is then applied to the channel features to generate a gating signal and modulate the spatial mixed features. Finally, inter-group information exchange is achieved through channel rearrangement and output. Layer normalization (LN) is then used to stabilize the feature distribution. The output of the enhanced grouped Mamba module fuses the results of the gating branch and the main branch through element-wise multiplication to obtain the output features of the enhanced grouped Mamba module.

5. The deep learning-based super-resolution method for industrial battery CT images according to claim 2, characterized in that, The hybrid channel attention feedforward module specifically includes: The input features are first fed into the MsConv hybrid convolutional unit, where depthwise convolutional branches with kernel sizes of 1×1, 3×3, and 5×5 are computed in parallel, and the outputs of each depthwise convolutional branch are concatenated to obtain the fused features. Then, the fused features are branched in two ways along the channel dimension. One branch is activated by GELU to obtain a gated branch, and the other branch serves as the main branch. The results of the two branches are multiplied element-wise. Finally, the channels are integrated through a PWConv layer and connected to the channel attention CA to obtain the output of the hybrid channel attention feedforward module.

6. The deep learning-based super-resolution method for industrial battery CT images according to claim 1, characterized in that, The frequency modulation module specifically includes: First, the input features are transformed to the frequency domain using FFT. The real and imaginary parts are concatenated along the channel dimension and then modeled using Dwconv, with residual connections used to preserve spectral information. Subsequently, the frequency domain features are fed into high-frequency and low-frequency branches for branch processing. The high-frequency branch extracts and modulates high-frequency related components through 3×1 and 1×3 convolutions, while the low-frequency branch extracts and modulates low-frequency related components through 3×3 convolutions. The outputs of the high-frequency and low-frequency branches are fused and fed into Pwconv and SiLU for integration. Finally, the output is back-mapped to the spatial domain using IFFT to obtain the output of the frequency modulation module.

Images