Microscopic image self-supervised denoising method and device, computer device and storage medium

Abstract

The application relates to a microscopic image self-supervised denoising method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring an original noisy microscopic image; a candidate set of random sampling starting points is constructed according to a preset down-sampling step; two mutually exclusive random sampling starting points are selected for an image block to be processed from the candidate set; non-overlapping slice sampling is performed on the image block to be processed extracted from the original noisy microscopic image according to the selected two random sampling starting points and the down-sampling step, so as to generate a set of training data pairs; a self-supervised denoising network is constructed and trained by using the generated training data pairs; and the microscopic image to be denoised is input into the trained network, so that a denoised microscopic image is obtained. The method can effectively break the spatial correlation of stripe artifacts, while maintaining the integrity of the periodic lattice structure, and realizes high-fidelity microscopic image denoising.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a self-supervised denoising method, apparatus, computer device, and storage medium for microscopic images. Background Technology

[0002] In the fields of computer vision and microscopic imaging data processing, especially in the denoising of electron microscope images (such as transmission electron microscopy (TEM) and scanning electron microscopy (SEM), self-supervised learning algorithms (such as Noise2Noise) have gradually become an important solution due to the difficulty in obtaining completely noise-free clean images as ground truth. Existing self-supervised methods typically employ fixed downsampling strategies when extracting training data from a single noisy image, such as sampling every other row and column starting from the top left corner, or using a simple random pixel discarding method to construct training image pairs.

[0003] However, in practical material characterization applications, the aforementioned traditional algorithms suffer from serious physical defects. First, they cannot effectively handle fringe artifacts with strong spatial correlation. Microscopic images often possess complex spatially correlated noise; for example, in scanning transmission electron microscopy, the electron beam's line-by-line scanning mechanism introduces highly directional scanning fringe artifacts. Traditional fixed downsampling methods completely preserve this regular noise, causing the network to mistakenly identify it as underlying physical structure and fail to filter it out. Second, when observing materials with periodic lattice crystals, fixed downsampling methods easily lead the network to learn fixed sampling biases, failing to reconstruct the true morphology of atoms. Summary of the Invention

[0004] Therefore, it is necessary to provide a self-supervised denoising method, apparatus, computer equipment, and storage medium for microscopic images that can effectively break the spatial correlation of noise and maintain the integrity of periodic structures, in order to address the above-mentioned technical problems.

[0005] A self-supervised denoising method for microscopic images, the method comprising:

[0006] Acquire the original noisy microscopic image and divide the original noisy microscopic image into multiple image blocks to be processed; Based on the preset downsampling step size, construct a candidate set of random sampling starting points; From the candidate set, two mutually exclusive random sampling starting points are selected for the image patch to be processed; Based on the two selected random sampling starting points and the downsampling step size, non-overlapping slice sampling is performed on the image block to generate a set of training data pairs. A self-supervised denoising network was constructed and trained using the generated training data. The microscopic image to be denoised is input into the trained self-supervised denoising network to obtain the denoised microscopic image.

[0007] In one embodiment, the method further includes: constructing a candidate set of offsets for the random sampling starting point based on the downsampling step size k. .

[0008] In one embodiment, the method further includes: randomly shuffling the elements in the candidate set and extracting the first two elements in the shuffled list as two independent random sampling starting points for the current image block.

[0009] In one embodiment, it further includes: based on the height of the input image patch and width And the downsampling step size k, calculate the target downsampling size: and ; Based on the first random sampling starting point Construct the left subgraph :

[0010] Based on the second random sampling starting point Construct the right subgraph :

[0011] Wherein, img is the image block.

[0012] In one embodiment, the method further includes: verifying whether the tensor shape of the generated training data pairs is equal to the target downsampling size, and retaining training data pairs that meet the size requirements.

[0013] In one embodiment, the method further includes: converting the original noisy microscopic image into a floating-point tensor suitable for deep neural networks, and constructing a multidimensional tensor input format using a dimension expansion operator.

[0014] In one embodiment, the self-supervised denoising network is further comprising: a deep convolutional neural network.

[0015] A self-supervised denoising device for microscopic images, the method comprising: The image acquisition module is used to acquire the original noisy microscopic image and divide the original noisy microscopic image into multiple image blocks to be processed; The candidate space construction module is used to construct a candidate set of random sampling starting points based on a preset downsampling step size; The starting point selection module is used to select two mutually exclusive random sampling starting points for the image block to be processed from the candidate set; The data pair generation module is used to perform non-overlapping slice sampling on the image patch according to the two selected random sampling starting points and the downsampling step size, and generate a set of training data pairs. The network training module is used to construct and train a self-supervised denoising network using the generated training data; The image denoising module is used to input the microscopic image to be denoised into the trained self-supervised denoising network to obtain the denoised microscopic image.

[0016] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps: Acquire the original noisy microscopic image and divide the original noisy microscopic image into multiple image blocks to be processed; Based on the preset downsampling step size, construct a candidate set of random sampling starting points; From the candidate set, two mutually exclusive random sampling starting points are selected for the image patch to be processed; Based on the two selected random sampling starting points and the downsampling step size, non-overlapping slice sampling is performed on the image block to generate a set of training data pairs. A self-supervised denoising network was constructed and trained using the generated training data. The microscopic image to be denoised is input into the trained self-supervised denoising network to obtain the denoised microscopic image.

[0017] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Acquire the original noisy microscopic image and divide the original noisy microscopic image into multiple image blocks to be processed; Based on the preset downsampling step size, construct a candidate set of random sampling starting points; From the candidate set, two mutually exclusive random sampling starting points are selected for the image patch to be processed; Based on the two selected random sampling starting points and the downsampling step size, non-overlapping slice sampling is performed on the image block to generate a set of training data pairs. A self-supervised denoising network was constructed and trained using the generated training data. The microscopic image to be denoised is input into the trained self-supervised denoising network to obtain the denoised microscopic image.

[0018] The aforementioned self-supervised denoising method, apparatus, computer equipment, and storage medium for microscopic images break the spatial correlation of noise under traditional fixed sampling methods by constructing a candidate set of random sampling starting points and selecting two mutually exclusive random starting points for non-overlapping sampling for each image block. This random starting point mechanism enables the deep network to accurately remove and suppress strongly directional stripe artifacts. Simultaneously, maintaining the fixity of the internal sampling grid during a single slicing process ensures the topological stability of the material's periodic structure, avoiding physical structure distortion, thereby achieving high-fidelity, high-quality automated denoising of microscopic images. Attached Figure Description

[0019] Figure 1 This is an illustration of an application scenario for a self-supervised denoising method for microscopic images in one embodiment. Figure 2 This is a flowchart illustrating a self-supervised denoising method for microscopic images in one embodiment; Figure 3 This is a denoising effect diagram of a self-supervised denoising method for microscopic images in one embodiment; wherein, Figure 3 (a) shows the effect of self-supervised denoising of the original experimental image under microscopic imaging; Figure 3 (b) shows the effect of self-supervised denoising of microscopic images using traditional fixed diagonal downsampling. Figure 3 (c) shows the effect of self-supervised denoising of microscopic images by downsampling from a random starting point; Figure 4 This is a schematic diagram of the downsampling process in another embodiment; wherein, Figure 4 (a) represents traditional fixed diagonal line sampling; Figure 4 (b) indicates downsampling from a random starting point; Figure 5 This is a generalization effect diagram of a self-supervised denoising method for microscopic images in one embodiment; wherein, Figure 5 (a) shows the generalization effect of self-supervised denoising of the original experimental image in microscopic images; Figure 5 (b) shows the generalization effect of traditional fixed diagonal downsampling for self-supervised denoising of microscopic images. Figure 5 (c) shows the generalization effect of self-supervised denoising of microscopic images by downsampling from a random starting point; Figure 6 This is a structural block diagram of a self-supervised denoising device for microscopic images in one embodiment; Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0021] The self-supervised denoising method for microscopic images provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. Terminal 102 can acquire or store raw noisy microscopic images and send the images to server 104. After receiving the image, server 104 executes the denoising method of this application, namely: acquiring the raw noisy microscopic image; constructing a candidate set of random sampling starting points according to a preset downsampling step size; selecting two mutually exclusive random sampling starting points for the image patch to be processed from the candidate set; performing non-overlapping slice sampling on the image according to the two selected starting points and step size to generate training data pairs; using these data to train a self-supervised denoising network; finally, inputting the image to be denoised into the network to obtain a denoised high-fidelity image, and returning the result to terminal 102 for display or storage. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices, and server 104 can be implemented using a standalone server or a server cluster composed of multiple servers.

[0022] In one embodiment, such as Figure 2 As shown, a self-supervised denoising method for microscopic images is provided, which is then applied to... Figure 1 Taking the server in the example, the following steps are included: Step 202: Obtain the original noisy microscopic image and divide the original noisy microscopic image into multiple image blocks to be processed.

[0023] Specifically, the server retrieves the raw microscopic image to be processed, which contains various types of noise (such as Gaussian noise, Poisson noise, and scanning fringe artifacts), from the terminal or local database. This image can be a grayscale or color image acquired by equipment such as a transmission electron microscope (TEM) or a scanning electron microscope (SEM).

[0024] Step 204: Construct a candidate set of random sampling starting points according to the preset downsampling step size.

[0025] Here, the downsampling step size is a preset parameter. For example, in a specific embodiment, the step size k can be set to 2. Constructing a candidate set of random sampling starting points is for the subsequent random selection of the starting position of sampling. For example, when the step size k=2, the candidate set C of offsets to the random sampling starting points can be defined as... This set contains four candidate points: (0,0), (0,1), (1,0), and (1,1). This candidate set provides all possible sampling start point offsets.

[0026] Step 206: Select two mutually exclusive random sampling starting points for the image patch to be processed from the candidate set.

[0027] To ensure that the generated data meets the noise independence requirement, two different and mutually exclusive random sampling starting points need to be selected for the same image patch. Specifically, the elements in the candidate set can first be randomly shuffled (e.g., using the `random.shuffle` function), and then the first two elements in the shuffled list can be extracted as the two independent random sampling starting points for the current image patch. This ensures that the input and target images generated for the same patch have random displacement differences in their spatial starting positions and are absolutely mutually exclusive.

[0028] Step 208: Based on the two selected random sampling starting points and the downsampling step size, perform non-overlapping slice sampling on the image blocks to generate a set of training data pairs.

[0029] This step utilizes the two random starting points selected in step 206, combined with a preset step size, to perform two independent, non-overlapping downsampling operations on the image patch. This sampling method ensures that the two sub-images used for training are completely non-overlapping at the pixel level, thus satisfying the premise of noise statistical independence in the Noise2Noise theory. Simultaneously, this non-periodic, skip-sampling effectively disrupts the noise continuity along the scanning direction.

[0030] Step 210: Construct and train a self-supervised denoising network using the generated training data.

[0031] The generated large amount of training data is input into a deep convolutional neural network. One subgraph is used as input and the other as the supervision target. The loss function is calculated and iteratively optimized until the model converges, and the trained denoising model is obtained.

[0032] Step 212: Input the microscopic image to be denoised into the trained self-supervised denoising network to obtain the denoised microscopic image.

[0033] The original or new noisy microscopic image is input into the trained model, and the model outputs a high-fidelity image that has removed the noise and preserved the original physical structure.

[0034] The aforementioned self-supervised denoising method for microscopic images constructs a candidate set of random sampling starting points and selects two mutually exclusive random starting points for non-overlapping sampling for each image patch, breaking the spatial correlation of noise under the traditional fixed sampling method. This random starting point mechanism enables the deep network to accurately remove and suppress strongly directional stripe artifacts. Simultaneously, maintaining the fixity of the internal sampling grid during a single slice ensures the topological stability of the material's periodic structure, avoiding physical structure distortion, thus achieving high-fidelity, high-quality automated denoising of microscopic images.

[0035] In one embodiment, step 206 may include: randomly shuffling the candidate list (e.g., using the random.shuffle function); and then extracting the first two elements from the shuffled list as the first random sampling starting points. Second random sampling start point Therefore, in each iteration, each image block obtains a pair of random, mutually exclusive starting coordinates, breaking the spatial regularity of noise at its source. This method mathematically disrupts the spatial correlation of fringe artifacts between adjacent pixels in microscopic images such as scanning transmission electron microscopy (STEM), forcibly satisfying the stringent theoretical conditions for the separation of incoherent noise.

[0036] In one embodiment, step 208 is implemented as follows: Based on the height of the input image patch and width And the downsampling step size k, calculate the target downsampling size: and Based on the first random sampling starting point Construct the left subgraph :

[0037] Based on the second random sampling starting point Construct the right subgraph :

[0038] Wherein, img is the image block.

[0039] The aforementioned design, with random starting points and a fixed internal grid, can both disrupt the continuity of noise and preserve the relative physical topological relationships between local pixels, thus effectively preventing the periodic lattice structure from being torn or distorted during downsampling. For materials with periodic lattices, the network extracts features of the same atomic column morphology through different random sampling starting points in different training iterations. This is equivalent to isotropic dense sampling of the two-dimensional spatial distribution of atoms at the sub-pixel level, constraining the network output to have rotation- and translation-invariant prediction results.

[0040] In one embodiment, to ensure the quality of the training data, a robustness verification step is included after generating a set of training data pairs. Specifically, an anomaly handling mechanism is introduced to verify the generated training data pairs ( and Is the tensor shape strictly equal to the expected target downsampling size? , The algorithm retains only regular samples that meet the size requirements and automatically removes invalid blocks that are incomplete in size due to edge cutting, thereby avoiding subsequent network operation crashes, ensuring the regularity of the training dataset, and finally outputting an aligned and regular self-supervised training dataset.

[0041] In one embodiment, after acquiring the original noisy microscopic image, an image preprocessing step is also included. Specifically, the original noisy microscopic image is converted into a floating-point tensor suitable for deep neural networks, such as float32 format. For the input 2D image, it is increased in dimensionality in the first dimension using a dimensionality expansion operator to construct a multidimensional tensor input format of [N, H, W], where N is the number of samples, H is the height, and W is the width, to ensure that the underlying data loader can correctly extract the image and provide correct dimensionality support for subsequent steps.

[0042] In one embodiment, the self-supervised denoising network is a deep convolutional neural network. This network structure may include an encoder-decoder structure, residual connections, etc., to effectively learn the mapping relationship from noisy images to clean images. Using the output regularized training dataset, this deep convolutional neural network denoising model is constructed. The generated mutually exclusive candidate data pairs are used as network inputs and constraint targets, respectively. The self-supervised loss function is calculated and iteratively optimized until the model converges, obtaining the trained denoising model.

[0043] In one specific embodiment, to verify the practical effect and versatility of the self-supervised denoising method for microscopic images proposed in this application under different imaging modalities, the applicant conducted comparative experiments under scanning transmission electron microscopy (STEM) and atomic force microscopy (AFM) characterization, respectively. The experimental results are as follows: Figure 3 and Figure 5 As shown. Figure 3and Figure 5 Each set of images was displayed: Figure 3 (a) shows the effect of self-supervised denoising of the original experimental image under microscopic imaging; Figure 3 (b) shows the effect of self-supervised denoising of microscopic images using traditional fixed diagonal downsampling. Figure 3 (c) shows the effect of self-supervised denoising of microscopic images by downsampling from a random starting point. Figure 5 (a) shows the generalization effect of self-supervised denoising of the original experimental image in microscopic images; Figure 5 (b) shows the generalization effect of traditional fixed diagonal downsampling for self-supervised denoising of microscopic images. Figure 5 (c) represents the generalization effect of self-supervised denoising of microscopic images by downsampling from a random starting point.

[0044] The experimental setup was as follows: Raw microscopic images of ReSe2, a two-dimensional transition metal chalcogenide with a periodic lattice structure, were acquired using STEM, as shown below. Figure 3 middle Figure 3 As shown in (a), the original image contains strong directional scanning stripe artifacts introduced by the electron beam progressive scan mechanism. While the atomic lattice structure in the image is clearly visible, it is affected by noise. In the experiment, the original image was processed using both the traditional fixed diagonal sampling self-supervised denoising method and the random starting point sampling self-supervised denoising method proposed in this application. The sampling methods are as follows: Figure 4 As shown, Figure 4 (a) represents traditional fixed diagonal line sampling; Figure 4 (b) indicates the random starting point downsampling of the present invention.

[0045] In another specific embodiment, to verify the generalization ability of this method, AFM was used to acquire raw high-resolution microscopic images of the two-dimensional transition metal chalcogenide MoS2 with a periodic lattice structure, such as... Figure 5 The first image in the table is shown. Unlike STEM, this original image contains strong directional scanning fringe artifacts introduced by the AFM probe's row-by-row scanning mechanism. The same denoising comparison experiment was performed on this high-resolution image. The results show that the random starting point sampling mechanism of this application can also effectively break the probe scanning fringe artifacts in the AFM image and faithfully restore the true surface periodic morphology of the raw material. This further demonstrates that the microscopic image denoising method of this invention is not limited to a single hardware device and possesses universality and excellent denoising generalization ability in various imaging modalities with fringe artifacts.

[0046] Specifically, in the processing of the method in this application, the downsampling step size k=2 is set, and a candidate set of random sampling starting points C = {(0,0), (0,1), (1,0), (1,1)} is constructed. For each image patch, the candidate set is randomly shuffled, and the first two elements are extracted as two mutually exclusive random sampling starting points, such as (0,0) and (1,1). Then, based on these two starting points, the original image is sliced ​​across rows and columns respectively to generate two non-overlapping training sub-images. For example, a sub-image is obtained by sampling with (0,0) as the starting point. The subgraph is obtained by sampling from (1,1). These two data pairs form a training data pair. A deep convolutional neural network is trained using a large number of such data pairs, and finally the original noisy image is input into the trained network to obtain the denoised result.

[0047] Experimental results are as follows Figure 3 and Figure 5 As shown: Figure 3 (b) and Figure 5 (b) shows that although the traditional fixed diagonal downsampling method can remove some random noise, significant scan stripe artifacts still remain in the image, and the edges of the atomic lattice structure show a certain degree of blurring and deformation, with some periodic structures failing to be fully restored. In contrast, Figure 3 (c) and Figure 5 (c) shows that after processing with the method of this application, all kinds of scanning stripe artifacts are significantly eliminated, and the overall image clarity is greatly improved. More importantly, the atomic lattice structure is restored with high fidelity, presenting a morphology that conforms to the laws of physical reality, with complete lattice periodicity, sharp edges, and no obvious deformation.

[0048] The above experimental results fully demonstrate the effectiveness of the technical solution presented in this application. The fundamental reason lies in the fact that this application, through a random starting point sampling mechanism, randomly jumps the coordinates of the sampling starting point each time training data pairs are constructed. This mathematically breaks the spatial correlation between adjacent pixels of fringe artifacts in STEM and AFM microscopic images, enabling the network to accurately remove and suppress this highly directional noise during the learning process. Simultaneously, for materials with periodic lattices, the network extracts features of the same atomic column morphology through different random sampling starting points in different training iteration cycles. This is equivalent to isotropic dense sampling of the two-dimensional spatial distribution of atoms at the sub-pixel level, constraining the network output to have rotation- and translation-invariant prediction results. This discards the directional bias caused by the fixed sampling grid, resulting in a final output lattice atomic morphology that more closely conforms to physical realities. Furthermore, the consistency of the internal sampling grid is maintained in the single-time subgraph construction, perfectly ensuring the topological stability of the material's periodic structure and avoiding physical structure distortion caused by excessive randomization.

[0049] It should be understood that, although Figure 2 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 2 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0050] In one embodiment, such as Figure 6 As shown, a self-supervised denoising device for microscopic images is provided, comprising: an image acquisition module 602, a candidate space construction module 604, a starting point selection module 606, a data pair generation module 608, a network training module 610, and an image denoising module 612, wherein: Image acquisition module 602 is used to acquire raw noisy microscopic images.

[0051] The candidate space construction module 604 is used to construct a candidate set of random sampling starting points according to a preset downsampling step size.

[0052] The starting point selection module 606 is used to select two mutually exclusive random sampling starting points for the image block to be processed from the candidate set.

[0053] The data pair generation module 608 is used to perform non-overlapping slice sampling on the image block to be processed according to the two selected random sampling starting points and the downsampling step size, and generate a set of training data pairs.

[0054] The network training module 610 is used to construct and train a self-supervised denoising network using the generated training data.

[0055] The image denoising module 612 is used to input the microscopic image to be denoised into the trained self-supervised denoising network to obtain the denoised microscopic image.

[0056] Specific limitations regarding the self-supervised denoising device for microscopic images can be found in the limitations of the self-supervised denoising method for microscopic images above, and will not be repeated here. Each module in the aforementioned self-supervised denoising device for microscopic images can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0057] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and the database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores raw and denoised microscopic image data. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a self-supervised denoising method for microscopic images.

[0058] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0059] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0060] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0061] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0062] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0063] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A self-supervised denoising method for microscopic images, characterized in that, The method includes: Acquire a raw, noisy microscopic image and divide the raw, noisy microscopic image into multiple image blocks to be processed; wherein the raw, noisy microscopic image contains fringe artifacts that are spatially correlated along the scanning direction; Based on the preset downsampling step size, construct a candidate set of random sampling starting points; From the candidate set, two mutually exclusive random sampling starting points are independently selected for each image block to be processed; Based on the two selected random sampling starting points and the downsampling step size, non-overlapping slice sampling is performed on the image patch to generate a set of training data pairs; wherein, during a single slice sampling process, the internal sampling grid remains fixed to maintain the relative topological relationship of the periodic structure; A self-supervised denoising network was constructed and trained using the generated training data. The microscopic image to be denoised is input into the trained self-supervised denoising network to obtain the denoised microscopic image.

2. The method according to claim 1, characterized in that, Based on the preset downsampling step size, a candidate set of random sampling starting points is constructed, including: Based on the downsampling step size k, construct a candidate set of offsets for the random sampling starting point. .

3. The method according to claim 1, characterized in that, From the candidate set, two mutually exclusive random sampling starting points are selected for the image patch to be processed, including: The elements in the candidate set are randomly shuffled, and the first two elements in the shuffled list are extracted as two independent random sampling starting points for the current image patch.

4. The method according to claim 1, characterized in that, Based on the two selected random sampling starting points and the downsampling step size, non-overlapping slice sampling is performed on the image patch, including: Based on the height of the input image patch and width And the downsampling step size k, calculate the target downsampling size: and ; Based on the first random sampling starting point Construct the left subgraph : Based on the second random sampling starting point Construct the right subgraph : Wherein, img is the image block.

5. The method according to claim 1, characterized in that, After generating a set of training data pairs, the following is also included: Verify whether the tensor shape of the generated training data pairs is equal to the target downsampling size, and retain the training data pairs that meet the size requirements.

6. The method according to claim 1, characterized in that, After acquiring the raw, noisy microscopic image, the process also includes: The original noisy microscopic image is converted into a floating-point tensor suitable for deep neural networks, and a multidimensional tensor input format is constructed through a dimension expansion operator.

7. The method according to any one of claims 1 to 6, characterized in that, The self-supervised denoising network is a deep convolutional neural network.

8. A self-supervised denoising device for microscopic images, characterized in that, The device includes: The image acquisition module is used to acquire the original noisy microscopic image and divide the original noisy microscopic image into multiple image blocks to be processed; wherein, the original noisy microscopic image contains fringe artifacts that have spatial correlation along the scanning direction; The candidate space construction module is used to construct a candidate set of random sampling starting points based on a preset downsampling step size; The starting point selection module is used to independently select two mutually exclusive random sampling starting points for each image block to be processed from the candidate set. The data pair generation module is used to perform non-overlapping slice sampling on the image patch according to the two selected random sampling starting points and the downsampling step size, and generate a set of training data pairs; wherein, during a single slice sampling process, the internal sampling grid remains fixed to maintain the relative topological relationship of the periodic structure; The network training module is used to construct and train a self-supervised denoising network using the generated training data; The image denoising module is used to input the microscopic image to be denoised into the trained self-supervised denoising network to obtain the denoised microscopic image.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images