Transmission electron microscope image denoising method and device based on block feature alignment

Abstract

The invention provides a transmission electron microscope image denoising method and device based on block feature alignment, and relates to the technical field of microscopic image characterization and analysis. The method comprises the following steps: acquiring an atomic microscopic video stream, and screening to obtain unpaired high-quality and low-quality data sets; respectively clustering the high-quality data set and the low-quality data set by using various clustering methods, obtaining a block containing an atomic region according to a clustering result and a voting mechanism, and carrying out unsupervised reconstruction training on the auto-encoder to obtain a pre-training model for high-quality and low-quality data reconstruction; and extracting features from the low-quality data set and the high-quality data set by using a pre-training model, projecting and aligning potential representations of the low-quality data set and the high-quality data set to a common potential space, and realizing low-quality image denoising. According to the transmission electron microscope image denoising method based on block feature alignment, the problems that in the prior art, atomic region recognition is not accurate, the denoising effect is limited, and a high and low quality image feature alignment mechanism is lacked can be solved, so that the image quality and the structure reduction precision are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of microscopic image characterization and analysis, and particularly to a transmission electron microscope image denoising method and device based on block feature alignment. Background Art

[0002] In the research of materials science, with the continuous update of image acquisition devices, the scale and quality of the generated datasets have increased exponentially. At present, the high-resolution mode of TEM (Transmission electron microscopy) can already obtain atomic-resolution images, which contain complex information such as atomic number, lattice parameter, lattice type, sample orientation, and defects. The information therein can be quantitatively explained by the positioning of atomic columns. However, the quality of experimental images generated by these instruments is often affected by random and deterministic distortions generated by the environment. Especially in low-dose or high-resolution imaging, the noise is more obvious. In addition, during the long imaging process, scanning distortions caused by device drift and environmental vibration may occur, resulting in image structure deformation; if the imaging speed is increased, the signal-to-noise ratio and imaging quality will decrease significantly, which is not conducive to dynamic observation with high spatial / temporal resolution. Therefore, the improvement of image quality is a key step to ensure accurate analysis, which can improve the signal-to-noise ratio of the image while retaining image details.

[0003] Traditional regularization image denoising models are divided into two categories: linear models and non-linear models. In the denoising process of TEM data, the retention of edge details is crucial. The patent with the publication number CN117830136A provides a processing method and system for transmission electron microscope images. The patent restores the collected original electron microscope images, calculates the pixel values of each pixel point in the original electron microscope images, maps the coordinates of the pixel points of the target image to the coordinates in the original electron microscope images, and restores the target image by the method of bilinear interpolation. The restored image is decomposed by wavelet transform to obtain high-frequency coefficients and low-frequency coefficients. The high-frequency coefficients are subjected to threshold quantization processing to obtain processed components, and wavelet reconstruction is performed based on the components and low-frequency coefficients to obtain the denoised image. However, this patent only performs image enhancement based on fixed rules and explicit algorithms, and it is difficult to adapt to the complex and diverse noise types and structural distortion problems in transmission electron microscope images. It lacks the modeling of the feature mapping relationship between high-quality and low-quality images, and cannot achieve more robust and generalized image denoising and structure restoration, and there are certain limitations in practical applications. Summary of the Invention

[0004] To solve the technical problems existing in the prior art, such as inaccurate identification of atomic regions, limited denoising effect, and lack of a high-low quality image feature alignment mechanism, an embodiment of the present invention provides a transmission electron microscope image denoising method and device based on block feature alignment. The technical solution is as follows:

[0005] On the one hand, a transmission electron microscope image denoising method based on block feature alignment is provided. This method is implemented by a transmission electron microscope image denoising device, and the method includes:

[0006] S1. Obtain an atomic microscopic video stream through a transmission electron microscope, and screen the images in the atomic microscopic video stream to obtain an unpaired high-quality data set and a low-quality data set.

[0007] S2. Use multiple clustering methods to cluster the high-quality data set and the low-quality data set respectively, and obtain the blocks containing atomic regions in the high-quality data set and the blocks containing atomic regions in the low-quality data set according to the clustering results and the voting mechanism.

[0008] S3. Perform unsupervised reconstruction training on the autoencoder according to the blocks containing atomic regions in the high-quality data set and the blocks containing atomic regions in the low-quality data set respectively, and obtain a pre-trained model for high-quality data reconstruction and a pre-trained model for low-quality data reconstruction.

[0009] Among them, the pre-trained model for high-quality data reconstruction includes a high-quality encoder and a high-quality decoder, and the pre-trained model for low-quality data reconstruction includes a low-quality encoder and a low-quality decoder.

[0010] S4. Obtain high-quality image block features according to the high-quality data set and the pre-trained model for high-quality data reconstruction, obtain low-quality image block features according to the low-quality data set and the pre-trained model for low-quality data reconstruction, project the high-quality image block features and the low-quality image block features into a common latent space, and use the joint loss function as a constraint to make the low-quality image block features gradually approach the high-quality image block features to obtain a trained low-quality encoder.

[0011] S5. Realize low-quality image denoising according to the trained low-quality encoder and the high-quality decoder.

[0012] Optionally, S2 includes:

[0013] S21. Crop the images in the high-quality data set and the low-quality data set respectively to obtain high-quality image blocks and low-quality image blocks.

[0014] S22. Cluster the high-quality image blocks using multiple clustering methods to obtain multiple clustering results of the high-quality image blocks, and fuse the multiple clustering results using the voting mechanism to obtain the blocks containing atomic regions in the high-quality data set.

[0015] S23. For low-quality image patches, use multiple clustering methods to perform clustering, obtain multiple clustering results of the low-quality image patches, and use a voting mechanism to fuse the multiple clustering results to obtain the patches containing atomic regions in the low-quality dataset.

[0016] Optionally, the multiple clustering methods include a clustering method based on the gray-scale features of the image, a clustering method based on the statistical features of the average gray value and gray standard deviation of the image, and a clustering method based on the frequency domain and structural statistical features of the image.

[0017] Optionally, in S3, the unsupervised reconstruction training of the autoencoder is respectively performed according to the patches containing atomic regions in the high-quality dataset and the patches containing atomic regions in the low-quality dataset to obtain a pre-trained model for high-quality data reconstruction and a pre-trained model for low-quality data reconstruction, including:

[0018] S31. Input the patches containing atomic regions in the high-quality dataset into a U-Net with an encoder-decoder structure, and perform unsupervised reconstruction training using the mean squared error loss to obtain a pre-trained model for high-quality data reconstruction.

[0019] S32. Input the patches containing atomic regions in the low-quality dataset into a U-Net with an encoder-decoder structure, and perform unsupervised reconstruction training using the mean squared error loss to obtain a pre-trained model for low-quality data reconstruction.

[0020] Optionally, S4 includes:

[0021] S41. Crop the images in the high-quality dataset and the low-quality dataset respectively to obtain high-quality image patches and low-quality image patches.

[0022] S42. Use the pre-trained model for high-quality data reconstruction to extract features from the high-quality image patches to obtain high-quality image patch features, and use the high-quality image patch features as the reference distribution.

[0023] S43. Use the pre-trained model for low-quality data reconstruction to extract features from the low-quality image patches to obtain low-quality image patch features.

[0024] S44. Freeze the high-quality encoder, use the low-quality encoder as the trainable model, project and map the high-quality image patch features and the low-quality image patch features into a common latent space, use the joint loss function to constrain the training of the network, and make the low-quality image patch features gradually approach the high-quality image patch features to obtain a trained low-quality encoder.

[0025] Optionally, the joint loss function is as shown in the following formula (1):

[0026] (1)

[0027] Where,

[0028] (2)

[0029] (3)

[0030] In the formula, represents the combined loss function, represents the weight coefficient occupied by represents the KL divergence loss function, represents the weight coefficient occupied by represents the mean squared error loss function, represents the number of channels, represents the variance of the output features of the encoder to be optimized, represents the variance of the output features of the frozen high-quality encoder, represents the mean of the output features of the encoder to be optimized, represents the mean of the output features of the frozen high-quality encoder, represents the total number of combinations of all channels and spatial positions, and represent the th features extracted by the low-quality encoder and the high-quality encoder in the same image region.

[0031] Optionally, S5 includes:

[0032] Obtain the low-quality transmission electron microscope image to be denoised, input it into the trained low-quality encoder to obtain output features, and reconstruct the output features according to the high-quality decoder to obtain the denoised image.

[0033] On the other hand, a transmission electron microscope image denoising device based on block feature alignment is provided. The device is applied to the transmission electron microscope image denoising method based on block feature alignment. The device includes:

[0034] An acquisition module, configured to acquire an atomic microscopy video stream through a transmission electron microscope, and screen the images in the atomic microscopy video stream to obtain an unpaired high-quality data set and a low-quality data set.

[0035] A clustering module, configured to use multiple clustering methods to cluster the high-quality data set and the low-quality data set respectively, and obtain the blocks containing atomic regions in the high-quality data set and the blocks containing atomic regions in the low-quality data set according to the clustering results and the voting mechanism.

[0036] A reconstruction module, configured to perform unsupervised reconstruction training on the autoencoder according to the blocks containing atomic regions in the high-quality data set and the blocks containing atomic regions in the low-quality data set respectively, to obtain a pre-trained model for high-quality data reconstruction and a pre-trained model for low-quality data reconstruction.

[0037] Among them, the pre-trained model for high-quality data reconstruction includes a high-quality encoder and a high-quality decoder, and the pre-trained model for low-quality data reconstruction includes a low-quality encoder and a low-quality decoder.

[0038] A training module, configured to obtain high-quality image patch features according to a high-quality data set and the pre-trained model for high-quality data reconstruction, obtain low-quality image patch features according to a low-quality data set and the pre-trained model for low-quality data reconstruction, project and map the high-quality image patch features and the low-quality image patch features into a common latent space, and use a joint loss function as a constraint to make the low-quality image patch features gradually approach the high-quality image patch features, so as to obtain a trained low-quality encoder.

[0039] An output module, configured to implement low-quality image denoising according to the trained low-quality encoder and the high-quality decoder.

[0040] Optionally, a clustering module, further configured to:

[0041] S21. Crop the images in the high-quality data set and the low-quality data set respectively to obtain high-quality image patches and low-quality image patches.

[0042] S22. For the high-quality image patches, perform clustering using multiple clustering methods to obtain multiple clustering results of the high-quality image patches, and use a voting mechanism to fuse the multiple clustering results to obtain the patches in the high-quality data set that contain atomic regions.

[0043] S23. For the low-quality image patches, perform clustering using multiple clustering methods to obtain multiple clustering results of the low-quality image patches, and use a voting mechanism to fuse the multiple clustering results to obtain the patches in the low-quality data set that contain atomic regions.

[0044] Optionally, the multiple clustering methods include a clustering method based on the gray-scale features of an image, a clustering method based on the statistical features of the average gray value and the standard deviation of the gray value of an image, and a clustering method based on the frequency domain and structural statistical features of an image.

[0045] Optionally, a reconstruction module, further configured to:

[0046] S31. Input the patches in the high-quality data set that contain atomic regions into a U-Net with an encoder-decoder structure, and perform unsupervised reconstruction training using the mean square error loss to obtain the pre-trained model for high-quality data reconstruction.

[0047] S32. Input the patches in the low-quality data set that contain atomic regions into a U-Net with an encoder-decoder structure, and perform unsupervised reconstruction training using the mean square error loss to obtain the pre-trained model for low-quality data reconstruction.

[0048] Optionally, the training module, further configured to:

[0049] S41. Crop the images in the high-quality dataset and the low-quality dataset respectively to obtain high-quality image patches and low-quality image patches.

[0050] S42. Extract features from the high-quality image patches using a pre-trained model for high-quality data reconstruction to obtain high-quality image patch features, and use the high-quality image patch features as the reference distribution.

[0051] S43. Extract features from the low-quality image patches using a pre-trained model for low-quality data reconstruction to obtain low-quality image patch features.

[0052] S44. Freeze the high-quality encoder, use the low-quality encoder as the trainable model, project and map the high-quality image patch features and the low-quality image patch features into a common latent space, and use a joint loss function to constrain the training of the network to make the low-quality image patch features gradually approach the high-quality image patch features to obtain a trained low-quality encoder.

[0053] Optionally, the joint loss function is as shown in the following formula (1):

[0054] (1)

[0055] Where

[0056] (2)

[0057] (3)

[0058] In the formula, represents the joint loss function, represents the weight coefficient occupied by represents the KL divergence loss function, represents the weight coefficient occupied by represents the mean squared error loss function, represents the number of channels, represents the variance of the output features of the encoder to be optimized, represents the variance of the output features of the frozen high-quality encoder, represents the mean of the output features of the encoder to be optimized, represents the mean of the output features of the frozen high-quality encoder, represents the total number of combinations of all channels and spatial positions, 、 represents the th feature extracted by the low-quality encoder and the high-quality encoder in the same image region.

[0059] Optionally, the output module is further used for:

[0060] Obtain a low-quality transmission electron microscope image to be denoised, input it into a trained low-quality encoder to obtain output features, and reconstruct the output features according to a high-quality decoder to obtain a denoised image.

[0061] On the other hand, a transmission electron microscope image denoising device is provided. The transmission electron microscope image denoising device includes: a processor; a memory, and computer-readable instructions are stored on the memory. When the computer-readable instructions are executed by the processor, any one of the methods in the above-mentioned transmission electron microscope image denoising method based on block feature alignment is implemented.

[0062] On the other hand, a computer-readable storage medium is provided. At least one instruction is stored in the storage medium, and the at least one instruction is loaded and executed by a processor to implement any one of the methods in the above-mentioned transmission electron microscope image denoising method based on block feature alignment.

[0063] The beneficial effects brought by the technical solutions provided by the embodiments of the present invention at least include:

[0064] In the present invention, an atomic microscopy video stream in an experiment is recorded by a transmission electron microscope CCD camera, the video is segmented by the number of frames, and high-quality and low-quality data are manually selected according to the image resolution, atomic clarity, and integrity of the crystal lattice structure; by using direct clustering, statistical feature clustering, and Fourier feature-based clustering methods for the high-quality and low-quality data sets respectively, the quality of the blocks of a single image is predicted using a global model and the results are fused through a voting mechanism; the selected low-quality and high-quality blocks are respectively used for unsupervised training of an autoencoder with a U-Net architecture to reconstruct the low-quality and high-quality data sets, and the pre-trained models are respectively saved to provide a basis for subsequent feature alignment; the extracted features are aligned in a common latent space to realize denoising of low-quality images. In this way, by combining traditional clustering with deep learning, the regions containing key atomic structures in the image can be effectively identified, and image enhancement under complex noise can be realized through the feature alignment mechanism. Compared with the existing rule-driven or single-channel reconstruction methods, the present invention is particularly suitable for dealing with the problem of signal-to-noise ratio degradation caused by high-resolution imaging, low dose, or sample perturbation in atomic-scale dynamic images, and provides a clearer and more credible basis for atomic-level high-resolution microscopic image analysis. Description of the Drawings

[0065] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0066] Figure 1It is a flowchart of a transmission electron microscope image denoising method based on block feature alignment provided by an embodiment of the present invention;

[0067] Figure 2 It is a detailed flowchart diagram of a transmission electron microscope image denoising method based on block feature alignment provided by an embodiment of the present invention;

[0068] Figure 3 It is a schematic diagram of the principle of block clustering provided by an embodiment of the present invention;

[0069] Figure 4 It is a schematic diagram of the result of block clustering provided by an embodiment of the present invention;

[0070] Figure 5 It is a schematic diagram of the principle of block feature alignment provided by an embodiment of the present invention;

[0071] Figure 6 It is a schematic diagram of the denoising result provided by an embodiment of the present invention;

[0072] Figure 7 It is a block diagram of a transmission electron microscope image denoising device based on block feature alignment provided by an embodiment of the present invention;

[0073] Figure 8 It is a schematic diagram of the structure of a transmission electron microscope image denoising device provided by an embodiment of the present invention. Detailed implementation manners

[0074] The following describes the technical solutions in the present invention with reference to the accompanying drawings.

[0075] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as "example" in the present invention should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Exactly speaking, the use of the word "example" is intended to present concepts in a specific manner. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one of the two.

[0076] In the embodiments of the present invention, "image" and "picture" can sometimes be used interchangeably. It should be noted that when their differences are not emphasized, their intended meanings are the same. "(of)", "corresponding", and "corresponding" can sometimes be used interchangeably. It should be noted that when their differences are not emphasized, their intended meanings are the same.

[0077] In the embodiments of the present invention, sometimes subscripts such as W1 may be written in a non-subscript form such as W1. When their differences are not emphasized, their intended meanings are the same.

[0078] To make the technical problems, technical solutions, and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments.

[0079] An embodiment of the present invention provides a method for denoising transmission electron microscope images based on block feature alignment. This method can be implemented by a transmission electron microscope image denoising device, which can be a terminal or a server. As Figure 1 shown in the flowchart of the method for denoising transmission electron microscope images based on block feature alignment, the processing flow of this method can include the following steps:

[0080] S1. Obtain an atomic microscopic video stream through a transmission electron microscope, screen the images in the atomic microscopic video stream, and obtain an unpaired high-quality data set and a low-quality data set.

[0081] In a feasible implementation manner, the method for denoising transmission electron microscope images based on block feature alignment of the present invention is as Figure 2 shown. Conduct experiments inside the transmission electron microscope, observe and record experimental phenomena by using the high-resolution mode, obtain an atomic-level high-resolution microscopic video stream, and perform frame segmentation on the obtained video. For the video after frame segmentation, respectively screen out a high-quality data set with high resolution, clear atomic points, and neat arrangement, and a low-quality data set with low resolution, blurred atomic points, and discontinuous lattice fringes.

[0082] Specifically, the recording device of the high-resolution transmission electron microscope is equipped with a slow-scan second CCD camera to continuously record the electronic signals generated during the experiment, convert the microscopic image optical signals received at the lower part of the fluorescent screen into digital signals, observe and record experimental phenomena, and obtain an experimental data set.

[0083] For the collected video data, it is necessary to keep the parameters, reference regions, clarity, and magnification consistent during the experiment, preliminarily screen the data therein to obtain key frames that meet the above conditions, and then the denoising of transmission electron microscopic images based on block feature alignment can be performed.

[0084] Furthermore, for the high-quality and low-quality data sets containing and pieces of data, the th piece of data is represented by , where .

[0085] S2. Use multiple clustering methods to cluster the high-quality data set and the low-quality data set respectively, and obtain the blocks containing atomic regions in the high-quality data set and the blocks containing atomic regions in the low-quality data set according to the clustering results and the voting mechanism.

[0086] In a feasible implementation manner, for two datasets, K-means clustering is respectively performed based on the original data, statistical features, and Fourier transform features. The quality of a single image block is predicted using a global model, and the results are fused through a voting mechanism to select the blocks containing the foreground of the atomic region.

[0087] Optionally, the above step S2 may include the following steps S21 - S23:

[0088] S21. Cut the images in the high-quality dataset and the low-quality dataset respectively to obtain high-quality image blocks and low-quality image blocks.

[0089] In a feasible implementation manner, each image in the dataset is cut into blocks of the same size, specifically 32×32 blocks.

[0090] S22. For the high-quality image blocks, multiple clustering methods are used for clustering to obtain multiple clustering results of the high-quality image blocks. The voting mechanism is used to fuse the multiple clustering results to obtain the blocks containing the atomic region in the high-quality dataset.

[0091] S23. For the low-quality image blocks, multiple clustering methods are used for clustering to obtain multiple clustering results of the low-quality image blocks. The voting mechanism is used to fuse the multiple clustering results to obtain the blocks containing the atomic region in the low-quality dataset.

[0092] Optionally, the multiple clustering methods include a clustering method based on the gray feature of the image, a clustering method based on the statistical features of the average gray value and gray standard deviation of the image, and a clustering method based on the frequency domain and structural statistical features of the image.

[0093] In a feasible implementation manner, as Figure 3 shown, the direct pixel value, statistical features, and Fourier transform features of each block are extracted, and they are classified as blocks belonging to the foreground region or the background region. The voting mechanism is used to statistically process the three K-means clustering methods to obtain the fused result and determine the blocks including the atomic region. Three K-means clustering models are trained on the entire dataset.

[0094] Specifically, each image is cut into sized blocks. For direct clustering, each block is represented as , and its feature vector is:

[0095] (1)

[0096] After flattening each image block into a one-dimensional vector, the features of all blocks are clustered, and K-means is used to cluster all the flattened vectors into two categories, and they are sorted according to the average pixel value after clustering.

[0097] For statistical feature clustering, the grayscale mean of each block is:

[0098] (2)

[0099] The grayscale standard deviation is:

[0100] (3)

[0101] Construct a two-dimensional feature vector with the grayscale mean and standard deviation. All blocks are divided into two categories through K-means clustering, and the means of all feature vectors in each category are statistically analyzed. The two clustering results are sorted in ascending order according to the comprehensive size of the means.

[0102] Furthermore, for the clustering method based on Fourier frequency domain features, perform frequency domain analysis on each image block using two-dimensional fast Fourier transform, and combine six feature dimensions including the number of significant peaks in the Fourier spectrum, the average distance of the peak from the frequency domain center, the proportion of high-frequency energy, the spectral information entropy, the periodic information, and the grayscale spatial variance into a six-dimensional feature vector. Perform clustering classification based on the above feature vector, and select the number of frequency domain peaks representing the periodic structural features as the evaluation basis. Sort the average value of this feature for each cluster in ascending order to determine the image quality level to which it belongs. In particular, in this embodiment, perform two-dimensional Fourier transform on the input image block as follows:

[0103] (4)

[0104] where represents the frequency component in the horizontal direction in the image frequency space, and represents the frequency component in the vertical direction in the image frequency space.

[0105] The obtained Fourier coefficients form the amplitude spectrum after amplitude operation , and further calculate the logarithmic amplitude spectrum:

[0106] (5)

[0107] After converting it into a one-dimensional signal , through local extreme detection and peak significance determination, set the significance threshold (such as ), to determine all indices that satisfy: . The significant peak count is used to reflect the structural information of the image block and is used as part of the frequency domain features in the subsequent clustering process.

[0108] Furthermore, the results of three K-means clustering are counted, the quality of blocks of a single image is predicted using the global model, and a voting mechanism is used to fuse the three methods to determine the blocks including the atomic regions, such as Figure 4 as shown

[0109] Furthermore, each block has three binary classification labels:

[0110] (6)

[0111] The final fusion result is:

[0112] (7)

[0113] S3. Respectively perform unsupervised reconstruction training on the autoencoder using the blocks containing atomic regions in the high-quality dataset and the blocks containing atomic regions in the low-quality dataset to obtain a pre-trained model for high-quality data reconstruction and a pre-trained model for low-quality data reconstruction.

[0114] Optionally, the above step S3 may include the following steps S31 - S32:

[0115] S31. Input the blocks containing atomic regions in the high-quality dataset into a U-Net with an encoder-decoder structure, and perform unsupervised reconstruction training using the mean squared error loss to obtain a pre-trained model for high-quality data reconstruction.

[0116] S32. Input the blocks containing atomic regions in the low-quality dataset into a U-Net with an encoder-decoder structure, and perform unsupervised reconstruction training using the mean squared error loss to obtain a pre-trained model for low-quality data reconstruction.

[0117] In a feasible implementation, the selected blocks of atomic regions are input into a U-Net with an encoder-decoder structure for unsupervised reconstruction training in the form of an autoencoder. Only the blocks containing atomic regions use the encoder to extract the latent features of the image, and the decoder restores and reconstructs the features, while the rest remains unchanged.

[0118] Furthermore, the mean squared error loss is used to measure the difference between each pixel point before and after reconstruction, and the reconstructed blocks are pieced back into the original image block by block. The mean squared error between the reconstructed image and the original image is compared to evaluate the effect of image restoration.

[0119] Save the pre-trained models for low-quality and high-quality dataset reconstructions respectively.

[0120] S4. Obtain high-quality image patch features based on the pre-trained model reconstructed from the high-quality dataset and high-quality data, obtain low-quality image patch features based on the pre-trained model reconstructed from the low-quality dataset and low-quality data, project and map the high-quality image patch features and the low-quality image patch features into a common latent space, use the joint loss function as a constraint, and make the low-quality image patch features gradually approach the high-quality image patch features to obtain a trained low-quality encoder.

[0121] In a feasible implementation manner, use the pre-trained model to extract the features of the high-quality and low-quality datasets respectively, project and map the features of the two into a common latent space, and train the network to optimize the model so that the low-quality features gradually align with the high-quality features to achieve low-quality image denoising, as Figure 5 shown.

[0122] Optionally, the above step S4 may include the following steps S41-S44:

[0123] S41. Crop the images in the high-quality dataset and the low-quality dataset respectively to obtain high-quality image patches and low-quality image patches.

[0124] In a feasible implementation manner, crop the low-quality and high-quality datasets into blocks and then load them into memory.

[0125] Specifically, crop the input size of the low-quality image into , divide the image into multiple blocks according to the size of , and send them into the neural network one by one as samples.

[0126] S42. Use the pre-trained model reconstructed from the high-quality data to extract features from the high-quality image patches to obtain high-quality image patch features, and use the high-quality image patch features as the reference distribution, denoted as .

[0127] S43. Use the pre-trained model reconstructed from the low-quality data to extract features from the low-quality image patches to obtain low-quality image patch features, denoted as .

[0128] In a feasible implementation manner, use the pre-trained model to extract the features of the blocks containing atomic regions in the low-quality and high-quality images respectively. The pre-trained model is considered to be able to better extract the latent features of the data.

[0129] S44. Freeze the high-quality encoder, use the low-quality encoder as the trainable model, project and map the high-quality image patch features and the low-quality image patch features into a common latent space, use the joint loss function to constrain the training of the network, and make the low-quality image patch features gradually approach the high-quality image patch features to obtain a trained low-quality encoder.

[0130] In a feasible implementation, the encoder part of the pre-trained model of the high-quality dataset is frozen, and the feature distribution obtained from the high-quality encoder is used as the reference target.

[0131] The encoder part of the pre-trained model of the low-quality dataset is used as the trainable model, and the goal is to align it to the output features of the high-quality encoder to improve the representation ability.

[0132] The feature distributions of the two are projected into a common latent space. In particular, the pre-trained model is considered to be able to better extract the latent features of the data.

[0133] Furthermore, in the alignment model stage, the network with the U-Net architecture is optimized, and the Kullback-Leibler divergence (K-L divergence) of the Gaussian distribution and the mean squared error loss are used as the metrics of the feature alignment mechanism to achieve the alignment reconstruction with the high quality as the reference.

[0134] To achieve the consistent alignment of the feature distributions, in the embodiments of the present invention, the Kullback-Leibier divergence (KL Divergence) based on the Gaussian hypothesis is introduced as the loss function on the encoder output feature map, and is defined as follows:

[0135] (8)

[0136] In the above formula, represents the variance of the output features of the encoder to be optimized, represents the variance of the output features of the frozen high-quality encoder, represents the mean of the output features of the encoder to be optimized, represents the mean of the output features of the frozen high-quality encoder, is the number of channels. This loss measures the degree of difference between the current network output feature distribution and the target distribution, so as to achieve the distribution alignment in the feature space.

[0137] Furthermore, to ensure the proximity of the encoder output features in the value range and at the same time suppress local distortion, in the embodiments of the present invention, the mean squared error (MSE) loss is also introduced as a supplementary metric, and is defined as follows:

[0138] (9)

[0139] Where and respectively represent the th features extracted by the low-quality and high-quality encoders in the same image region, represents the total number of combinations of all channels and spatial positions.

[0140] Finally, the example of the present invention uses a weighted combination of KL divergence and mean square error as the joint optimization objective to simultaneously guide the alignment of feature distributions and the approximation of feature values. The overall optimization objective function is as follows:

[0141] (10)

[0142] Furthermore, the decoder of the high-quality block is used to reconstruct the features of the low-quality encoder after feature alignment to obtain the denoised image.

[0143] The present invention adopts a method combining traditional clustering and deep learning, which can effectively identify the regions in the image containing key atomic structures and achieve image enhancement under complex noise through the feature alignment mechanism. Compared with the existing rule-driven or single-channel reconstruction methods, the present invention is particularly suitable for dealing with the problem of signal-to-noise ratio reduction in atomic-scale dynamic images caused by high-resolution imaging, low dose or sample perturbation, providing a clearer and more reliable basis for atomic-level high-resolution microscopic image analysis.

[0144] S5. Implement low-quality image denoising according to the trained low-quality encoder and high-quality decoder.

[0145] Specifically, obtain the low-quality transmission electron microscope image to be denoised, input it into the trained low-quality encoder to obtain the output features, and reconstruct the output features according to the high-quality decoder to obtain the denoised image.

[0146] In the embodiment of the present invention, the atomic microscopic video stream in the experiment is recorded by a transmission electron microscope CCD camera, the video is segmented according to the number of frames, and high-quality and low-quality data are manually selected according to the image resolution, atomic clarity and integrity of the lattice structure; by using direct clustering, statistical feature clustering and Fourier feature-based clustering methods for the high-quality and low-quality data sets respectively, the quality of the blocks of a single image is predicted using a global model and the results are fused through a voting mechanism; the selected low-quality and high-quality blocks are used to perform unsupervised training on an autoencoder with a U-Net architecture to reconstruct the low-quality and high-quality data sets, and the pre-trained models are saved respectively to provide a basis for subsequent feature alignment; the extracted features are aligned in a common latent space to achieve denoising of the low-quality image. In this way, by adopting a method combining traditional clustering and deep learning, the regions in the image containing key atomic structures can be effectively identified, and image enhancement under complex noise can be achieved through the feature alignment mechanism. Compared with the existing rule-driven or single-channel reconstruction methods, the present invention is particularly suitable for dealing with the problem of signal-to-noise ratio reduction in atomic-scale dynamic images caused by high-resolution imaging, low dose or sample perturbation, providing a clearer and more reliable basis for atomic-level high-resolution microscopic image analysis.

[0147] Figure 7It is a block diagram of a transmission electron microscope image denoising device based on block feature alignment shown according to an exemplary embodiment. This device is used for the transmission electron microscope image denoising method based on block feature alignment. Refer to Figure 7 , this device includes an acquisition module 310, a clustering module 320, a reconstruction module 330, a training module 340, and an output module 350. Among them:

[0148] The acquisition module 310 is used to obtain an atomic microscopy video stream through a transmission electron microscope, and screen the images in the atomic microscopy video stream to obtain an unpaired high-quality data set and a low-quality data set.

[0149] The clustering module 320 is used to cluster the high-quality data set and the low-quality data set respectively using multiple clustering methods, and obtain the blocks containing atomic regions in the high-quality data set and the blocks containing atomic regions in the low-quality data set according to the clustering results and the voting mechanism.

[0150] The reconstruction module 330 is used to perform unsupervised reconstruction training on the autoencoder according to the blocks containing atomic regions in the high-quality data set and the blocks containing atomic regions in the low-quality data set, and obtain a pre-trained model for high-quality data reconstruction and a pre-trained model for low-quality data reconstruction.

[0151] Among them, the pre-trained model for high-quality data reconstruction includes a high-quality encoder and a high-quality decoder, and the pre-trained model for low-quality data reconstruction includes a low-quality encoder and a low-quality decoder.

[0152] The training module 340 is used to obtain high-quality image block features according to the high-quality data set and the pre-trained model for high-quality data reconstruction, obtain low-quality image block features according to the low-quality data set and the pre-trained model for low-quality data reconstruction, project and map the high-quality image block features and the low-quality image block features into a common latent space, and use a joint loss function as a constraint to make the low-quality image block features gradually approach the high-quality image block features, and obtain a trained low-quality encoder.

[0153] The output module 350 is used to realize low-quality image denoising according to the trained low-quality encoder and the high-quality decoder.

[0154] In the embodiments of the present invention, an atomic microscopy video stream in an experiment is recorded by a transmission electron microscope CCD camera, the video is segmented by the number of frames, and high-quality and low-quality data are manually selected according to the image resolution, atomic clarity, and integrity of the lattice structure; by using direct clustering, statistical feature clustering, and Fourier feature-based clustering methods for the high-quality and low-quality data sets respectively, the quality of the blocks of a single image is predicted using a global model and the results are fused through a voting mechanism; the selected low-quality and high-quality blocks are used for unsupervised training of an autoencoder with a U-Net architecture to reconstruct the low-quality and high-quality data sets, and the pre-trained models are saved respectively to provide a basis for subsequent feature alignment; the extracted features are aligned in a common latent space to denoise the low-quality images. In this way, by combining traditional clustering with deep learning, the regions containing key atomic structures in the images can be effectively identified, and image enhancement under complex noise can be achieved through the feature alignment mechanism. Compared with the existing rule-driven or single-channel reconstruction methods, the present invention is particularly suitable for dealing with the problem of signal-to-noise ratio degradation caused by high-resolution imaging, low dose, or sample perturbation in atomic-scale dynamic images, providing a clearer and more reliable basis for atomic-level high-resolution microscopy image analysis.

[0155] Figure 8 FIG. 4 is a schematic structural diagram of a transmission electron microscope image denoising device provided by an embodiment of the present invention, as Figure 8 shown, the transmission electron microscope image denoising device may include the above-mentioned Figure 7 transmission electron microscope image denoising device based on block feature alignment shown in FIG. Optionally, the transmission electron microscope image denoising device 410 may include a first processor 2001.

[0156] Optionally, the transmission electron microscope image denoising device 410 may further include a memory 2002 and a transceiver 2003.

[0157] Wherein, the first processor 2001 is connected to the memory 2002 and the transceiver 2003, such as through a communication bus.

[0158] Next, in conjunction with Figure 8 FIG., each component of the transmission electron microscope image denoising device 410 will be specifically introduced:

[0159] Among them, the first processor 2001 is the control center of the transmission electron microscope image denoising device 410, which can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 is one or more central processing units (CPUs), or can be an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention, such as: one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs).

[0160] Optionally, the first processor 2001 can execute various functions of the transmission electron microscope image denoising device 410 by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0161] In a specific implementation, as an embodiment, the first processor 2001 can include one or more CPUs, such as Figure 8 the CPU0 and CPU1 shown in

[0162] In a specific implementation, as an embodiment, the transmission electron microscope image denoising device 410 can also include multiple processors, such as Figure 8 the first processor 2001 and the second processor 2004 shown in. Each of these processors can be a single-core processor (single-CPU) or a multi-core processor (multi-CPU). Here, the processor can refer to one or more devices, circuits, and / or processing cores for processing data (such as computer program instructions).

[0163] Among them, the memory 2002 is used to store software programs for implementing the solution of the present invention and is controlled by the first processor 2001 for execution. The specific implementation method can refer to the above method embodiments and will not be elaborated here.

[0164] Optionally, the memory 2002 may be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or may also be an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM), or other optical disc storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media, or other magnetic storage devices, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory 2002 may be integrated with the first processor 2001 or may exist independently and be coupled to the first processor 2001 through an interface circuit ( Figure 8 not shown) of the transmission electron microscope image denoising device 410. The embodiments of the present invention do not make specific limitations on this.

[0165] The transceiver 2003 is used to communicate with a network device or with a terminal device.

[0166] Optionally, the transceiver 2003 may include a receiver and a transmitter ( Figure 8 not shown separately). Among them, the receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0167] Optionally, the transceiver 2003 may be integrated with the first processor 2001 or may exist independently and be coupled to the first processor 2001 through an interface circuit ( Figure 8 not shown) of the transmission electron microscope image denoising device 410. The embodiments of the present invention do not make specific limitations on this.

[0168] It should be noted that Figure 8 the structure of the transmission electron microscope image denoising device 410 shown does not constitute a limitation on the router. The actual knowledge structure recognition device may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0169] In addition, the technical effects of the transmission electron microscope image denoising device 410 may refer to the technical effects of the transmission electron microscope image denoising method based on block feature alignment described in the above method embodiments, and will not be elaborated here.

[0170] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0171] It should also be understood that the memory in the embodiments of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM) or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM) and direct rambus RAM (DR RAM).

[0172] The above embodiments can be implemented in whole or in part by software, hardware (such as circuits), firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present invention are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wired (such as infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that contains one or more collections of available media. The available media can be magnetic media (such as floppy disks, hard disks, magnetic tapes), optical media (such as DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0173] It should be understood that the term "and / or" in this document is merely a description of the association relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. Here, A and B can be singular or plural. Additionally, the character " / " in this document generally represents an "or" relationship between the associated objects before and after, but it may also represent an "and / or" relationship, which can be understood specifically by referring to the context before and after.

[0174] In the present invention, "at least one" means one or more, and "a plurality" means two or more. "At least one of the following" or its similar expressions refer to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, or c can represent: a, b, c, a - b, a - c, b - c, or a - b - c, where a, b, and c can be single or multiple.

[0175] It should be understood that in various embodiments of the present invention, the magnitudes of the sequence numbers of the above processes do not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

[0176] Those of ordinary skill in the art will realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented in electronic hardware, or in a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled professionals can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0177] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the devices, apparatuses, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0178] In several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there can be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0179] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they can be located in one place, or can be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0180] In addition, the functional units in various embodiments of the present invention can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit.

[0181] When the above-mentioned functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or a part of this technical solution can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0182] As described above, the above are only specific implementation manners of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A transmission electron microscope image denoising method based on block feature alignment, characterized in that The method includes: S1. Obtain an atomic microscopy video stream through a transmission electron microscope, screen the images in the atomic microscopy video stream to obtain an unpaired high-quality data set and a low-quality data set; S2. Use multiple clustering methods to cluster the high-quality data set and the low-quality data set respectively, and obtain the blocks containing atomic regions in the high-quality data set and the blocks containing atomic regions in the low-quality data set according to the clustering results and the voting mechanism; S3. Perform unsupervised reconstruction training on the autoencoder according to the blocks containing atomic regions in the high-quality data set and the blocks containing atomic regions in the low-quality data set respectively to obtain a pre-trained model for high-quality data reconstruction and a pre-trained model for low-quality data reconstruction; Among them, the pre-trained model for high-quality data reconstruction includes a high-quality encoder and a high-quality decoder, and the pre-trained model for low-quality data reconstruction includes a low-quality encoder and a low-quality decoder; S4. Obtain high-quality image block features according to the high-quality data set and the pre-trained model for high-quality data reconstruction, obtain low-quality image block features according to the low-quality data set and the pre-trained model for low-quality data reconstruction, project the high-quality image block features and the low-quality image block features into a common latent space, use the joint loss function as a constraint, and make the low-quality image block features gradually approach the high-quality image block features to obtain a trained low-quality encoder; S5. Implement low-quality image denoising according to the trained low-quality encoder and the high-quality decoder.

2. The method for denoising a transmission electron microscope image based on block feature alignment according to claim 1, wherein The S2 includes: S21. Crop the images in the high-quality data set and the low-quality data set respectively to obtain high-quality image blocks and low-quality image blocks; S22. Cluster the high-quality image blocks using multiple clustering methods to obtain multiple clustering results of the high-quality image blocks, and fuse the multiple clustering results using the voting mechanism to obtain the blocks containing atomic regions in the high-quality data set; S23. Cluster the low-quality image blocks using multiple clustering methods to obtain multiple clustering results of the low-quality image blocks, and fuse the multiple clustering results using the voting mechanism to obtain the blocks containing atomic regions in the low-quality data set.

3. The method for denoising a transmission electron microscope image based on block feature alignment according to claim 2, wherein The multiple clustering methods include a clustering method based on the gray-scale features of the image, a clustering method based on the statistical features of the average gray value and the standard deviation of the gray scale of the image, and a clustering method based on the frequency domain and structural statistical features of the image.

4. The transmission electron microscope image denoising method based on block feature alignment according to claim 1, wherein The performing unsupervised reconstruction training on the autoencoder according to the blocks containing atomic regions in the high-quality data set and the blocks containing atomic regions in the low-quality data set respectively in S3 to obtain a pre-trained model for high-quality data reconstruction and a pre-trained model for low-quality data reconstruction includes: S31. Input the blocks containing atomic regions in the high-quality data set into a U-Net with an encoder-decoder structure, and perform unsupervised reconstruction training using the mean square error loss to obtain a pre-trained model for high-quality data reconstruction; S32. Input the blocks containing atomic regions in the low-quality data set into a U-Net with an encoder-decoder structure, and perform unsupervised reconstruction training using the mean square error loss to obtain a pre-trained model for low-quality data reconstruction.

5. The method for denoising a transmission electron microscope image based on block feature alignment according to claim 1, wherein The S4 includes: S41. Crop the images in the high-quality dataset and the low-quality dataset respectively to obtain high-quality image patches and low-quality image patches; S42. Extract features from the high-quality image patches using a pre-trained model for high-quality data reconstruction to obtain high-quality image patch features, and use the high-quality image patch features as the reference distribution; S43. Extract features from the low-quality image patches using a pre-trained model for low-quality data reconstruction to obtain low-quality image patch features; S44. Freeze the high-quality encoder, use the low-quality encoder as the trainable model, project and map the high-quality image patch features and the low-quality image patch features into a common latent space, and use a joint loss function to constrain the training of the network to make the low-quality image patch features gradually approach the high-quality image patch features to obtain a trained low-quality encoder.

6. The method for denoising a transmission electron microscope image based on block feature alignment according to claim 5, wherein The joint loss function is shown in the following formula (1): （1） Wherein, （2） （3） In the formula, represents the combined loss function, represents the weight coefficient occupied by represents the KL divergence loss function, represents the weight coefficient occupied by represents the mean squared error loss function, represents the number of channels, represents the variance of the output features of the encoder to be optimized, represents the variance of the output features of the frozen high-quality encoder, represents the mean of the output features of the encoder to be optimized, represents the mean of the output features of the frozen high-quality encoder, represents the total number of combinations of all channels and spatial positions, , represent the th features extracted by the low-quality encoder and the high-quality encoder in the same image region.

7. The method for denoising transmission electron microscope images based on block feature alignment according to claim 1, wherein The S5 includes: Obtain a low-quality transmission electron microscope image to be denoised, input it into the trained low-quality encoder to obtain output features, and reconstruct the output features according to the high-quality decoder to obtain a denoised image.

8. A transmission electron microscope image denoising device based on block feature alignment, the transmission electron microscope image denoising device based on block feature alignment is used to implement the transmission electron microscope image denoising method based on block feature alignment according to any one of claims 1-7, characterized in that, The device includes: An acquisition module, configured to acquire an atomic microscopy video stream through a transmission electron microscope, and screen the images in the atomic microscopy video stream to obtain an unpaired high-quality dataset and low-quality dataset; A clustering module, configured to use multiple clustering methods to cluster the high-quality dataset and the low-quality dataset respectively, and obtain the blocks containing atomic regions in the high-quality dataset and the blocks containing atomic regions in the low-quality dataset according to the clustering results and a voting mechanism; A reconstruction module, configured to perform unsupervised reconstruction training on the autoencoder according to the blocks containing atomic regions in the high-quality dataset and the blocks containing atomic regions in the low-quality dataset respectively to obtain a pre-trained model for high-quality data reconstruction and a pre-trained model for low-quality data reconstruction; Wherein, the pre-trained model for high-quality data reconstruction includes a high-quality encoder and a high-quality decoder, and the pre-trained model for low-quality data reconstruction includes a low-quality encoder and a low-quality decoder; A training module, configured to obtain high-quality image patch features according to the high-quality dataset and the pre-trained model for high-quality data reconstruction, obtain low-quality image patch features according to the low-quality dataset and the pre-trained model for low-quality data reconstruction, project and map the high-quality image patch features and the low-quality image patch features into a common latent space, and use a joint loss function as a constraint to make the low-quality image patch features gradually approach the high-quality image patch features to obtain a trained low-quality encoder; An output module, configured to perform low-quality image denoising according to the trained low-quality encoder and the high-quality decoder.

9. A transmission electron microscope image denoising device, characterized in that The transmission electron microscope image denoising device includes: A processor; A memory, on which computer-readable instructions are stored, and when the computer-readable instructions are executed by the processor, the method described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that, Program code is stored in the computer-readable storage medium, and the program code can be called by the processor to execute the method described in any one of claims 1 to 7.

Images