Image denoising model training and image denoising method, image denoising model training device, equipment and medium

Abstract

The application discloses an image denoising model training and image denoising method, an image denoising model training device, equipment and a medium, and is applied to the field of image denoising. A sample image is decomposed into a low-frequency subband and a high-frequency subband through discrete wavelet transform. The high-frequency subband is input into a high-frequency subband hierarchical aggregation module to obtain an enhanced high-frequency subband. The low-frequency subband is input into a low-frequency subband space-frequency domain hybrid enhancement module to obtain an enhanced low-frequency subband. Discrete wavelet inverse transform is performed based on the enhanced high-frequency subband and the enhanced low-frequency subband to obtain an enhanced image. Model loss is determined based on the enhanced image. Model parameters are updated based on the model loss. The training of the image denoising model is completed until the image denoising model is obtained. A to-be-denoised image is input into the image denoising model to obtain a denoised enhanced image. The application decomposes an image through discrete wavelet transform, and reduces image noise through high-frequency subband hierarchical aggregation and low-frequency subband space-frequency domain hybrid enhancement, thereby improving defect detection precision and efficiency.

Description

Technical Field

[0001] This invention relates to the field of image denoising, and particularly to an image denoising model training method, an image denoising method, an image denoising model training device, an electronic device, and a computer-readable storage medium. Background Technology

[0002] In the quality control stage of lithium battery and nuclear material manufacturing, image denoising is a crucial step in improving defect detection accuracy. Lithium battery surfaces need to identify micron-level scratches, while nuclear materials require the detection of micron- or even submicron-level cracks. However, on-site images are often affected by Gaussian, salt-and-pepper, or mixed noise. Traditional methods (such as Gaussian filtering) rely on manual features and have limited performance. While convolutional neural network-based methods improve denoising quality, their high computational complexity and limited receptive field make them unsuitable for the real-time requirements of industrial production lines. Wavelet domain methods decompose images into low-frequency and high-frequency subbands using discrete wavelet transform and combine them with convolutional networks for lossless downsampling. However, high-frequency subbands lack multi-scale interaction, low-frequency subbands suffer from insufficient global structure optimization, and their robustness to complex noise is limited. Selective state-space models excel in visual tasks due to their linear complexity and long-range dependency modeling capabilities, but they have not yet been specifically optimized for wavelet subbands. Summary of the Invention

[0003] The purpose of this invention is to provide an image denoising model training and image denoising method, image denoising model training device, equipment and medium, which are applied in the field of image denoising. The method uses discrete wavelet transform to decompose the image and enhances it by high-frequency subband layer aggregation and low-frequency subband space-frequency domain hybridization to reduce image noise and thus improve the accuracy and efficiency of defect detection.

[0004] To address the aforementioned technical problems, this invention provides an image denoising model training method, comprising: The sample image is input into the image denoising model, and the sample image is decomposed into low-frequency sub-band and high-frequency sub-band through discrete wavelet transform; The high-frequency subband is downsampled, pre-state space model calculated and upsampled by the high-frequency subband hierarchical aggregation module. The upsampling result is fused with the high-frequency subband feature, and the fusion result is used to calculate the post-state space model to obtain the enhanced high-frequency subband. The low-frequency subband is subjected to Fourier transform, state space model calculation and inverse Fourier transform by the low-frequency subband spatial-frequency domain hybrid enhancement module. The spatial features obtained by convolving the inverse Fourier transform result with the low-frequency subband are fused to obtain the enhanced low-frequency subband. An enhanced image is obtained by performing discrete wavelet inverse transform based on the enhanced high-frequency subband and the enhanced low-frequency subband. The model loss is determined based on the enhanced image, and the model parameters are updated based on the model loss until the trained image denoising model is obtained.

[0005] Optionally, the discrete wavelet transform is a first-order discrete wavelet transform; the low-frequency subband includes: LL; the high-frequency subband includes: LH, HL and HH.

[0006] Optionally, the high-frequency subband is downsampled, pre-state space model calculated, and upsampled using a high-frequency subband hierarchical aggregation module. The upsampling result is then fused with the high-frequency subband features, and the fused result is used for post-state space model calculation to obtain an enhanced high-frequency subband, including: The LH is input to the high-frequency subband layer aggregation module 1, the HL is input to the high-frequency subband layer aggregation module 2, and the HH is input to the high-frequency subband layer aggregation module 3. The high-frequency subband is downsampled, pre-state space model calculated, and upsampled by the high-frequency subband hierarchical aggregation module 1, the high-frequency subband hierarchical aggregation module 2, and the high-frequency subband hierarchical aggregation module 3, respectively. The upsampling result is fused with the high-frequency subband feature, and the fused result is used to perform post-state space model calculation to obtain the enhanced high-frequency subband.

[0007] Optionally, the state-space model computation is based on a selective state-space model; the selective state-space model includes: block partitioning and serialization, sequence computation and deserialization.

[0008] Optionally, the convolution is a 3×3 convolution; the upsampling and downsampling factors are both 2x.

[0009] To address the aforementioned technical problems, this invention provides an image denoising method, comprising: Obtain the image to be denoised, and input the image to be denoised into the image denoising model to obtain the denoised enhanced image; The image denoising model is a model trained according to the image denoising model training method described above.

[0010] Optionally, the method further includes: The enhanced image is input into the defect detection model to obtain the output defect detection result; The image to be denoised can be of any type, either a lithium battery image or a nuclear material image; the defect detection model is a model trained based on any type of lithium battery image and / or nuclear material image.

[0011] To address the aforementioned technical problems, the present invention provides an image denoising model training device, comprising: The first module is used to input the sample image into the image denoising model and decompose the sample image into a low-frequency sub-band and a high-frequency sub-band through discrete wavelet transform. The second module is used to perform downsampling, pre-state space model calculation and upsampling on the high-frequency sub-band through the high-frequency sub-band hierarchical aggregation module, fuse the upsampling result with the high-frequency sub-band feature, and perform post-state space model calculation on the fusion result to obtain the enhanced high-frequency sub-band. The third module is used to perform Fourier transform, state space model calculation and inverse Fourier transform on the low-frequency sub-band through the low-frequency sub-band spatial-frequency domain hybrid enhancement module, and to perform feature fusion by convolving the inverse Fourier transform result with the spatial features obtained by the low-frequency sub-band to obtain the enhanced low-frequency sub-band. The fourth module is used to perform discrete wavelet inverse transform based on the enhanced high-frequency subband and the enhanced low-frequency subband to obtain an enhanced image, determine the model loss based on the enhanced image, update the model parameters based on the model loss, until the trained image denoising model is obtained.

[0012] To solve the above-mentioned technical problems, the present invention provides an electronic device, comprising: Memory, used to store computer programs; A processor is configured to implement the image denoising model training method or the image denoising method described above when executing the computer program.

[0013] To address the aforementioned technical problems, the present invention provides a computer-readable storage medium storing computer-executable instructions. When these computer-executable instructions are executed by a processor, they implement the image denoising model training method or the image denoising method described above.

[0014] As can be seen, this invention inputs the sample image into the image denoising model, decomposes the sample image into low-frequency sub-bands and high-frequency sub-bands through discrete wavelet transform; the high-frequency sub-band is downsampled, pre-state space model calculated, and upsampled through a high-frequency sub-band hierarchical aggregation module, the upsampling result is fused with the high-frequency sub-band features, and the fused result is then used for post-state space model calculation to obtain an enhanced high-frequency sub-band; the low-frequency sub-band is Fourier transformed, state space model calculated, and inverse Fourier transformed through a low-frequency sub-band spatial-frequency domain hybrid enhancement module, the inverse Fourier transform result is convolved with the low-frequency sub-band to obtain spatial features, and then fused to obtain an enhanced low-frequency sub-band; an enhanced image is obtained by performing inverse discrete wavelet transform based on the enhanced high-frequency sub-band and enhanced low-frequency sub-band, the model loss is determined based on the enhanced image, and the model parameters are updated based on the model loss until a trained image denoising model is obtained; the image to be denoised is acquired, and the image to be denoised is input into the image denoising model to obtain the denoised enhanced image.

[0015] This invention uses discrete wavelet transform to decompose images and enhances them by layering high-frequency subbands and spatial-frequency domain mixing of low-frequency subbands, thereby reducing image noise and improving the accuracy and efficiency of defect detection. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] Figure 1 A flowchart illustrating an image denoising model training method provided in an embodiment of the present invention; Figure 2 This is a structural example diagram of a high-frequency subband hierarchical aggregation module provided in an embodiment of the present invention; Figure 3 This is a structural example diagram of a low-frequency sub-band spatial-frequency domain hybrid enhancement module provided in an embodiment of the present invention; Figure 4 This is a structural example diagram of an image denoising model provided in an embodiment of the present invention; Figure 5 This is a structural block diagram of an image denoising model training device provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The following combination Figure 1 , Figure 1 A flowchart of an image denoising model training method provided in an embodiment of the present invention, the method may include: S101: Input the sample image into the image denoising model, and decompose the sample image into low-frequency sub-band and high-frequency sub-band through discrete wavelet transform.

[0020] In this embodiment, a sample image can be input into an image denoising model, and the enhanced image can be obtained after processing by the image denoising model. The sample image in this embodiment can be an image with pre-added noise.

[0021] After inputting the model, this embodiment can decompose the sample image into low-frequency sub-bands and high-frequency sub-bands using discrete wavelet transform.

[0022] To improve computational efficiency, the discrete wavelet transform in this embodiment can be a first-order discrete wavelet transform. The first-order discrete wavelet transform decomposes the sample image into a low-frequency sub-band and a high-frequency sub-band. The low-frequency sub-band includes LL (LowLow, low-frequency), and the high-frequency sub-band includes LH (LowHigh, low-frequency), HL (HighLow, high-frequency), and HH (HighHigh, high-frequency).

[0023] S102: The high-frequency subband is downsampled, pre-state space model is calculated and upsampled through the high-frequency subband hierarchical aggregation module. The upsampling result is fused with the high-frequency subband feature and the fusion result is calculated by the post-state space model to obtain the enhanced high-frequency subband.

[0024] In this embodiment, each high-frequency subband can be input into its corresponding high-frequency subband hierarchical aggregation module for processing. For example, for LH, HL and HH obtained by first-level discrete wavelet transform decomposition, LH can be input into high-frequency subband hierarchical aggregation module 1, HL into high-frequency subband hierarchical aggregation module 2, and HH into high-frequency subband hierarchical aggregation module 3.

[0025] The high-frequency subband is downsampled, pre-state space model calculated, and upsampled by the high-frequency subband hierarchical aggregation module 1, high-frequency subband hierarchical aggregation module 2, and high-frequency subband hierarchical aggregation module 3, respectively. The upsampling result is fused with the high-frequency subband feature (such as pixel-by-pixel addition), and the fusion result is used to calculate the post-state space model to obtain the enhanced high-frequency subband.

[0026] In this embodiment, the structure of the high-frequency subband hierarchical aggregation module can be as follows: Figure 2 As shown, for any high-frequency subband hierarchical aggregation module, the input high-frequency subband will flow to path 1 and path 2 respectively.

[0027] At path 1, the high-frequency subband is downsampled, the pre-state space model is calculated, and upsampling is performed sequentially. After upsampling, feature fusion is performed with the data from path 2 (i.e., the initial high-frequency subband). After feature fusion, the post-state space model is calculated.

[0028] In this embodiment, the pre-state space model calculation and the post-state space model calculation are two different state space model calculations. The state space model calculation in this embodiment may include: block partitioning and serialization, sequence calculation, and deserialization.

[0029] like Figure 2As shown, the state-space model computation first divides the input image data into blocks, for example, image block 1, image block 2, image block 3, and image block 4. Then, the block-wise data is serialized, converting the 2D image data into 1D sequence data. After serialization, sequence computation is performed to extract global information, followed by deserialization, transforming the 1D sequence data back into 2D image data.

[0030] In this embodiment, the upsampling is 2x upsampling. The specific form of downsampling can be interpolation downsampling or pooling downsampling, etc. The specific form of upsampling can be bilinear interpolation or nearest neighbor interpolation, etc. There is no limitation here, and it can be set according to the actual application.

[0031] For high-frequency subbands, multi-scale hierarchical processing is used, combined with long-range dependency modeling of state space model, to capture cross-scale texture and edge information, thereby enhancing the ability to recover complex noisy images.

[0032] S103: The low-frequency subband is subjected to Fourier transform, state space model calculation and inverse Fourier transform through the low-frequency subband spatial-frequency domain hybrid enhancement module. The spatial features obtained by convolving the inverse Fourier transform result with the low-frequency subband are fused to obtain the enhanced low-frequency subband.

[0033] The input to the low-frequency subband space-frequency domain hybrid enhancement module is the low-frequency subband of the discrete wavelet transform, i.e., LL, such as... Figure 3 As shown, in this embodiment, the low-frequency subband spatial-frequency domain hybrid enhancement can include two branches: a spatial branch and a frequency domain branch.

[0034] The spatial features obtained by convolving the low-frequency subband through the spatial branch of the low-frequency subband spatial-frequency domain hybrid enhancement module are obtained. The spatial branch consists of only one lightweight convolutional layer, which can be a 3×3 convolution. The convolution type is not limited to standard convolution, group convolution, or depth convolution. This branch focuses on extracting local spatial features, such as details of local textures, edges, or smooth regions.

[0035] In the frequency domain branch, the input data is first transformed from the time domain to the frequency domain through Fourier transform (such as fast Fourier transform), then the state space model is calculated, and then the inverse Fourier transform is performed to restore the frequency domain signal to the time domain signal.

[0036] The output data of the spatial branch and the frequency domain branch need to be fused (e.g., pixel-by-pixel) to obtain the final output.

[0037] For the low-frequency sub-band, the system processes local features in the spatial domain and global features in the frequency domain in parallel. It uses a state-space model to enhance the consistency of the global structure. Frequency domain analysis can capture the global frequency distribution, while the state-space model calculation enhances the long-term dependencies of global features through serialization, thus making up for the limitations of convolution in long-range modeling and improving the denoising effect in smooth regions.

[0038] The state space model calculation in the low-frequency subband spatial-frequency domain hybrid enhancement module is the same as that in the high-frequency subband hierarchical aggregation module, both including image segmentation and serialization, sequence calculation and deserialization.

[0039] In this embodiment, the state-space model calculation can be based on a selective state-space model, which is based on state equations and observation equations, as shown in the following formula: h t =Ah t-1 +Bx t ; y t =Ch t ; In the formula, t is the discrete time, A, B, and C represent the state transition weight matrix, input weight matrix, and output weight matrix, respectively, and x t The input representing the current time, h t h represents the current state. t-1 Represents the previous state, y t This represents the current time.

[0040] S104: Based on the enhanced high-frequency subband and the enhanced low-frequency subband, perform discrete wavelet inverse transform to obtain the enhanced image, determine the model loss based on the enhanced image, update the model parameters based on the model loss, until the trained image denoising model is obtained.

[0041] The number of paths in the image denoising model is related to the wavelet decomposition level and can be set based on the actual application. An example of the image denoising model structure in this embodiment is as follows: Figure 4 As shown, when wavelet decomposition is performed into a first-order discrete wavelet transform, the image denoising model can contain four paths, of which paths 1, 2 and 3 perform feature extraction calculations for high-frequency subbands (LH, HL and HH); path 4 performs feature extraction calculations for low-frequency subbands (LL).

[0042] After feature extraction of all subbands is completed, four updated enhanced subbands (enhanced LH, enhanced HL, enhanced HH, and enhanced LL) are obtained. Finally, the four enhanced subband images are combined into one image by discrete inverse wavelet transform to obtain the final denoised image.

[0043] This embodiment can calculate the model loss based on the denoised image output by the model and the sample image, update the model parameters based on the model loss, and perform iterative training through the sample image set until a trained image denoising model is obtained.

[0044] This embodiment utilizes the long-range dependency modeling capability of the state space model through a high-frequency subband hierarchical aggregation module to accurately restore the texture and edge details of complex defects such as scratches on the lithium battery surface and cracks in the core material. At the same time, the low-frequency subband spatial-frequency domain hybrid enhancement module optimizes the global background consistency, ensuring that smooth areas are clear and natural. Compared with traditional convolutional network methods, this significantly improves the visual quality of the denoised image and the reliability of defect detection.

[0045] Furthermore, by employing the linear complexity calculation characteristics of the state-space model and combining it with a concise architecture of only one-level wavelet decomposition, the computational resource requirements are significantly reduced. This enables the rapid processing of high-resolution images, meeting the high efficiency requirements of real-time quality inspection in lithium battery and nuclear material production, and is particularly suitable for the strict time constraints of automated production lines.

[0046] Furthermore, by employing differentiated processing strategies for high-frequency and low-frequency subbands, it effectively addresses various interference types such as Gaussian noise, salt-and-pepper noise, and mixed noise, significantly improving denoising stability in complex industrial environments (such as low light and dust interference), and ensuring robustness and applicability in scenarios such as lithium battery surface inspection and nuclear material crack identification.

[0047] Based on the above embodiments, the present invention uses discrete wavelet transform to decompose images and enhances them by high-frequency subband hierarchical aggregation and low-frequency subband spatial-frequency domain hybridization, thereby reducing image noise and improving defect detection accuracy and efficiency.

[0048] The following is an image noise reduction method provided by the present invention, which may include: Obtain the image to be denoised, and input the image to be denoised into the image denoising model to obtain the denoised and enhanced image.

[0049] The image denoising model is a model trained according to the image denoising model training method.

[0050] This embodiment can acquire an image to be denoised. The image to be denoised can be an image of the object to be inspected taken by an industrial camera. This embodiment does not limit the specific type of the image to be denoised. The image to be denoised can be a lithium battery image and / or a nuclear material image.

[0051] Furthermore, this embodiment also provides a defect detection scheme to complement image noise reduction, so as to realize the entire industrial product defect detection process.

[0052] This embodiment can obtain the output defect detection results from the input of enhanced images to the defect detection model. This embodiment does not limit the specific architecture and training method of the defect detection model, and can be set according to the actual application. In this embodiment, the defect detection model can be a model trained based on lithium battery images and / or nuclear material images.

[0053] To address the dual requirements of accuracy and efficiency in image denoising for scenarios such as lithium battery production and nuclear material manufacturing, this embodiment enhances the recovery capability of details such as scratches on the lithium battery surface and cracks in nuclear materials through high-frequency subband layered aggregation, optimizes the global background consistency of low-frequency subbands, and improves the quality of denoised images. By utilizing the linear complexity of the selective state-space model and the first-level decomposition strategy, computational costs are reduced to meet the efficiency requirements of real-time industrial quality inspection. Through differentiated subband processing, it adapts to Gaussian, salt-and-pepper, and mixed noise, improving the stability and applicability of denoising in complex industrial environments.

[0054] The following combination Figure 5 , Figure 5 This is a structural block diagram of an image denoising model training device provided in an embodiment of the present invention. The device may include: The first module 100 is used to input the sample image into the image denoising model and decompose the sample image into a low-frequency sub-band and a high-frequency sub-band through discrete wavelet transform. The second module 200 is used to downsample, calculate the pre-state space model and upsample the high-frequency subband through the high-frequency subband hierarchical aggregation module, fuse the upsampling result with the high-frequency subband feature, and calculate the enhanced high-frequency subband by performing a post-state space model on the fusion result. The third module 300 is used to perform Fourier transform, state space model calculation and inverse Fourier transform on the low frequency subband through the low frequency subband spatial-frequency domain hybrid enhancement module, and to fuse the spatial features obtained by convolving the inverse Fourier transform result with the low frequency subband to obtain the enhanced low frequency subband. The fourth module 400 is used to perform discrete wavelet inverse transform based on the enhanced high-frequency subband and the enhanced low-frequency subband to obtain the enhanced image, determine the model loss based on the enhanced image, update the model parameters based on the model loss, until the trained image denoising model is obtained.

[0055] Based on the above embodiments, the present invention uses discrete wavelet transform to decompose images and enhances them by high-frequency subband hierarchical aggregation and low-frequency subband spatial-frequency domain hybridization, thereby reducing image noise and improving defect detection accuracy and efficiency.

[0056] Based on the above embodiments, the discrete wavelet transform is a first-order discrete wavelet transform; the low-frequency subband includes: LL; the high-frequency subband includes: LH, HL and HH.

[0057] Based on the above embodiments, the second module 200 may include: The first unit is used to input LH into the high-frequency subband layer aggregation module 1, HL into the high-frequency subband layer aggregation module 2, and HH into the high-frequency subband layer aggregation module 3. The second unit is used to perform downsampling, pre-state space model calculation and upsampling on the high-frequency subband through the high-frequency subband hierarchical aggregation module 1, high-frequency subband hierarchical aggregation module 2 and high-frequency subband hierarchical aggregation module 3 respectively, fuse the upsampling results with the high-frequency subband features, and perform post-state space model calculation on the fusion results to obtain the enhanced high-frequency subband.

[0058] Based on the above embodiments, the state space model calculation is based on the selective state space model; the selective state space model includes: block partitioning and serialization, sequence calculation and deserialization.

[0059] Based on the above embodiments, the convolution is a 3×3 convolution; the upsampling and downsampling factors are both 2 times.

[0060] The following is an image noise reduction device provided by the present invention, which may include: The fifth module is used to acquire the image to be denoised, and input the image to be denoised into the image denoising model to obtain the denoised enhanced image; Among them, the image denoising model is the model trained by the image denoising model training device.

[0061] Based on the above embodiments, the device may further include: The sixth module is used to input the enhanced image into the defect detection model and obtain the output defect detection result; The image to be denoised can be of any type, either a lithium battery image or a nuclear material image; the defect detection model is a model trained based on any type of lithium battery image or nuclear material image.

[0062] Based on the above embodiments, the present invention also provides an electronic device, which may include a memory and a processor. The memory stores a computer program, and when the processor calls the computer program in the memory, it can implement the steps provided in the above embodiments. Of course, the device may also include various necessary network interfaces, a power supply, and other components.

[0063] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by an execution terminal or processor, can implement the method provided in the embodiments of the present invention; the storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0064] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0065] In this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, without necessarily requiring or implying 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 limitation, 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.

Claims

1. A method for training an image denoising model, characterized in that, include: The sample image is input into the image denoising model, and the sample image is decomposed into low-frequency sub-band and high-frequency sub-band through discrete wavelet transform; The high-frequency subband is downsampled, pre-state space model calculated and upsampled by the high-frequency subband hierarchical aggregation module. The upsampling result is fused with the high-frequency subband feature, and the fusion result is used to calculate the post-state space model to obtain the enhanced high-frequency subband. The low-frequency subband is subjected to Fourier transform, state space model calculation and inverse Fourier transform by the low-frequency subband spatial-frequency domain hybrid enhancement module. The spatial features obtained by convolving the inverse Fourier transform result with the low-frequency subband are fused to obtain the enhanced low-frequency subband. An enhanced image is obtained by performing discrete wavelet inverse transform based on the enhanced high-frequency subband and the enhanced low-frequency subband. The model loss is determined based on the enhanced image, and the model parameters are updated based on the model loss until the trained image denoising model is obtained.

2. The image denoising model training method according to claim 1, characterized in that, The discrete wavelet transform is a first-order discrete wavelet transform; The low-frequency subband includes: LL; the high-frequency subband includes: LH, HL and HH.

3. The image denoising model training method according to claim 2, characterized in that, The high-frequency subband is downsampled, pre-state space model calculated, and upsampled using a high-frequency subband hierarchical aggregation module. The upsampling result is then fused with the high-frequency subband features, and the fused result is further calculated using a post-state space model to obtain an enhanced high-frequency subband. This process includes: The LH is input to the high-frequency subband layer aggregation module 1, the HL is input to the high-frequency subband layer aggregation module 2, and the HH is input to the high-frequency subband layer aggregation module 3. The high-frequency subband is downsampled, pre-state space model calculated, and upsampled by the high-frequency subband hierarchical aggregation module 1, the high-frequency subband hierarchical aggregation module 2, and the high-frequency subband hierarchical aggregation module 3, respectively. The upsampling result is fused with the high-frequency subband feature, and the fused result is used to perform post-state space model calculation to obtain the enhanced high-frequency subband.

4. The image denoising model training method according to claim 1, characterized in that, The state-space model computation is based on the selective state-space model; the selective state-space model includes: block partitioning and serialization, sequence computation and deserialization.

5. The image denoising model training method according to claim 1, characterized in that, The convolution is a 3×3 convolution; the upsampling and downsampling factors are both 2.

6. An image denoising method, characterized in that, include: Obtain the image to be denoised, and input the image to be denoised into the image denoising model to obtain the denoised enhanced image; The image denoising model is a model trained by the image denoising model training method according to any one of claims 1 to 5.

7. The image denoising method according to claim 6, characterized in that, Also includes: The enhanced image is input into the defect detection model to obtain the output defect detection result; The image to be denoised can be of any type, either a lithium battery image or a nuclear material image. The defect detection model is a model trained on any type of lithium battery image and nuclear material image.

8. An image denoising model training device, characterized in that, include: The first module is used to input the sample image into the image denoising model and decompose the sample image into a low-frequency sub-band and a high-frequency sub-band through discrete wavelet transform. The second module is used to perform downsampling, pre-state space model calculation and upsampling on the high-frequency sub-band through the high-frequency sub-band hierarchical aggregation module, fuse the upsampling result with the high-frequency sub-band feature, and perform post-state space model calculation on the fusion result to obtain the enhanced high-frequency sub-band. The third module is used to perform Fourier transform, state space model calculation and inverse Fourier transform on the low-frequency sub-band through the low-frequency sub-band spatial-frequency domain hybrid enhancement module, and to perform feature fusion by convolving the inverse Fourier transform result with the spatial features obtained by the low-frequency sub-band to obtain the enhanced low-frequency sub-band. The fourth module is used to perform discrete wavelet inverse transform based on the enhanced high-frequency subband and the enhanced low-frequency subband to obtain an enhanced image, determine the model loss based on the enhanced image, update the model parameters based on the model loss, until the trained image denoising model is obtained.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the image denoising model training method as described in any one of claims 1 to 5, or the image denoising method as described in any one of claims 6 or 7, when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, implement the image denoising model training method as described in any one of claims 1 to 5, or the image denoising method as described in any one of claims 6 or 7.

Images