Microscopic imaging method and system based on deep learning

Abstract

The invention discloses a microscopic imaging method and system based on deep learning, and relates to the technical field of optical microscopic imaging. The method comprises the following steps: acquiring a to-be-imaged image set obtained through sub-pixel sampling; the method comprises the following steps: constructing an image reconstruction model, training the image reconstruction model, carrying out variable sequence information fusion processing on a sample image set by a generator, learning shallow information in an image to obtain a fused image sequence, carrying out deep feature extraction mapping on the fused image sequence based on a multi-dimensional attention mechanism, and carrying out deep feature extraction mapping on the fused image sequence; and finally, carrying out feedback correction on image data by utilizing a residual image, carrying out discrimination feedback by a discriminator according to a result of the generator, and adjusting parameters of an image reconstruction model. And processing the image set by using the optimized image reconstruction model to obtain a reconstructed image. According to the super-resolution reconstruction method disclosed by the invention, high-quality and high-efficiency super-resolution reconstruction of a low-resolution image is realized by utilizing a plurality of images obtained by sub-pixel sampling of imaging equipment.

Description

Technical Field

[0001] The present invention relates to the field of optical microscopic imaging technology, and in particular to a microscopic imaging method and system based on deep learning. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Optical microscopy has become an indispensable tool for exploring the microscopic world. The rapid development of microscopic imaging has not only promoted the advancement of science and technology, but also put forward higher requirements on the performance of optical microscopy. As a key carrier for characterizing microstructures, high-resolution microscopic images can accurately present important information such as the spatial configuration, subtle features, and boundary contours of samples. Due to the inherent physical properties of microscopic imaging systems and the limitations of spatiotemporal resolution in the digital image sampling process, there is inevitable information degradation in images, resulting in low-quality phenomena such as optical blur, aliasing, artifacts, and insufficient detail information, which seriously affects the effectiveness and reliability of microscopic images, and thus affects the accuracy and efficiency of micro-nano detection / measurement.

[0004] With the rapid development of emerging technologies such as artificial intelligence and machine learning, the application of deep learning technology has greatly promoted the development of computational microscopy technology. Convolutional neural networks have excellent detail representation capabilities to effectively enhance the detail features of reconstructed images and have been widely used in the field of image super-resolution.

[0005] In order to further improve the resolution and accuracy of the image, the concept of sub-pixel is proposed. Sub-pixel is the smallest unit of the imaging surface of the area array camera with pixels as the smallest unit, and the distance between the pixel centers ranges from a few to more than ten microns. Sub-pixel is a smaller pixel unit that exists between two physical pixels and is obtained by calculation.

[0006] However, the current deep learning methods do not deeply mine sub-pixel displacement information, and single image reconstruction is mostly used. The input analysis of multi-angle image sets is not comprehensive enough. Therefore, the reconstructed images obtained cannot meet the requirements of the micro-nano field for high-quality and efficient microscopic imaging. Summary of the invention

[0007] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a microscopic imaging method and system based on deep learning, design a microscopic imaging device that uses a precision motion platform to acquire multiple images, and propose a microscopic image super-resolution method based on deep learning. By utilizing the multiple images obtained by sub-pixel sampling of the imaging device, high-quality and efficient super-resolution reconstruction of low-resolution images is achieved.

[0008] In order to achieve the above object, the present invention is implemented through the following technical solutions: A first aspect of the present invention provides a microscopic imaging method based on deep learning, comprising the following steps: Acquire an image set to be imaged obtained by sub-pixel sampling; Construct an image reconstruction model and train it to obtain an optimized image reconstruction model, wherein the image reconstruction model includes a generator and a discriminator. The generator first performs variable sequence information fusion processing on the sample image set, learns the shallow information in the image, and obtains a fused image sequence. Then, based on the multi-dimensional attention mechanism, the deep-level features of the fused image sequence are extracted and mapped to obtain a deep feature sequence. Finally, the residual map is used to feedback and correct the image data. The discriminator performs discriminative feedback based on the generator results and adjusts the image reconstruction model parameters. The image set is processed using the optimized image reconstruction model to obtain a reconstructed image.

[0009] Furthermore, an image set is acquired through an imaging device, which includes a microscopic imaging unit, a composite light source unit, a precision motion platform and a controller. The microscopic imaging unit is used to capture image data, and is composed of a CMOS camera and a microscope. The composite light source units are installed on both sides of the precision motion platform to control the light source and improve the light source and environment during the operation of the equipment. The precision motion platform performs precise displacement. The precision motion platform is controlled by the controller and performs image acquisition after moving to the specified position. The controller connects the computer and the precision motion platform to realize the transmission of motion instructions and the feedback of information.

[0010] Furthermore, the specific steps of obtaining the image set to be imaged obtained by sub-pixel sampling are: The imaging device receives the motion command and moves the precision motion platform to the specified position, collects the image of the specified position, and then uses the specified position as the reference point and collects the image of the adjacent position of the reference point according to the template of the adjacent position search. At this time, the precision motion platform has moved to the adjacent position of the reference point, and obtains the displacement information from the current position to other adjacent positions. The images of each position are collected by the precision motion platform to form a sequence image set.

[0011] Furthermore, the specific steps of performing variable sequence information fusion processing on the image set are: The image set is screened and classified according to the relative positions in the pixel space to obtain image sequences of different groups; Shallow feature learning is performed on different groups of image sequences, and each group of shallow feature learning results is concatenated to obtain a fused image sequence.

[0012] Furthermore, when extracting and mapping deep features, a residual dense block with batch normalization layer removed is introduced into the multi-dimensional attention mechanism to solve the problems of loss of detail feature information and gradient disappearance or explosion.

[0013] Furthermore, the specific steps of using the residual map to perform feedback correction on the image data are as follows: Perform multi-view upsampling feedback operations on the low-resolution image through sub-pixel convolution to obtain an upsampled image; Furthermore, the specific steps for the discriminator to perform discriminative feedback based on the generator results are: The convolution feature extraction unit is used to gradually refine and amplify the feature information of the image; At the end of the network, the features are further processed through two fully connected layers, and finally the authenticity probability of the image is output through the Sigmoid function; Calculate the loss function and generate discriminative feedback based on the calculation results.

[0014] A second aspect of the present invention provides a microscopic imaging system based on deep learning, comprising: An image acquisition module is configured to acquire an image set to be imaged obtained by sub-pixel sampling; The reconstruction model training module is configured to construct an image reconstruction model and train the image reconstruction model to obtain an optimized image reconstruction model, wherein the image reconstruction model includes a generator and a discriminator. The generator first performs variable sequence information fusion processing on the sample image set, learns the shallow information in the image, and obtains a fused image sequence. Then, based on the multi-dimensional attention mechanism, the deep-level features of the fused image sequence are extracted and mapped to obtain a deep feature sequence. Finally, the residual map is used to feedback and correct the image data. The discriminator performs discrimination feedback based on the generator result and adjusts the image reconstruction model parameters. The image reconstruction module is configured to process the image set using the optimized image reconstruction model to obtain a reconstructed image.

[0015] The third aspect of the present invention provides a medium having a program stored thereon, which, when executed by a processor, implements the steps in the deep learning-based microscopy imaging method as described in the first aspect of the present invention.

[0016] The fourth aspect of the present invention provides a device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the deep learning-based microscopy imaging method as described in the first aspect of the present invention are implemented.

[0017] One or more of the above technical solutions have the following beneficial effects: The present invention discloses a microscopic imaging method and system based on deep learning, and proposes a new microscopic imaging device. The device uses the precision motion of a precision motion platform to achieve multi-view sampling of an object, obtain multiple images of the object and sub-pixel displacement information, and provide a more solid and comprehensive data foundation for the next step of the image super-resolution algorithm. The device and algorithm have certain advantages in the application of microscopic size measurement. After line width and chamfer measurement experimental analysis, the device and algorithm have advantages in sub-micron measurement accuracy.

[0018] The present invention proposes a new method for achieving image super-resolution, designs the structures of the generator and discriminator modules, integrates the variable sequence fusion module, the multi-dimensional attention mechanism module, and the residual map feedback module into a unified closed loop. Unlike traditional interpolation methods and other deep learning methods, the algorithm is not limited to the perspective of single image reconstruction, and can perform super-resolution reconstruction of sequence images. The clarity and texture details of the reconstructed image have been improved, and a higher value has been achieved in the reconstructed image quality evaluation value.

[0019] In the field of image super-resolution, most existing methods reconstruct the angle of a single image, which has certain limitations in imaging details and texture. The present invention obtains sufficient multi-angle images and sub-pixel displacement information by using the designed imaging device and imaging method, and inputs this information into the deep learning algorithm to complete super-resolution image reconstruction. This invention opens up a new direction in the field of image super-resolution.

[0020] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0022] Figure 1 This is a structural diagram of an imaging device according to the first embodiment of the present invention; Figure 2 This is a structural diagram of a microscopic image super-resolution reconstruction algorithm based on deep learning in this embodiment; Figure 3 This is a structural diagram of a variable sequence fusion module in this embodiment; Figure 4 This is a structural diagram of the multi-dimensional attention mechanism module of the first embodiment of the present invention; Figure 5 This is a schematic diagram of the residual feedback process of the first embodiment of the present invention; Figure 6 4 is a structural diagram of a residual graph feedback module according to the first embodiment of the present invention. DETAILED DESCRIPTION

[0023] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0024] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "include" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or their combinations; Embodiment 1: Embodiment 1 of the present invention provides a microscopic imaging method based on deep learning, which mainly includes two parts. The first part is to use a microscopic imaging device to obtain multiple images and sub-pixel displacement information, and the second part is to input the captured low-resolution images and displacement information into a neural network algorithm to perform super-resolution reconstruction of the image. The specific steps include: Step 1: Obtain an image set to be imaged by sub-pixel sampling.

[0025] In a specific implementation, this embodiment acquires an image set through an imaging device, such as Figure 1 As shown, the imaging device includes a microscopic imaging unit, a composite light source unit, a precision motion platform and a controller. The microscopic imaging unit is used to capture image data and is composed of a CMOS camera and a microscope. The composite light source unit is installed on both sides of the precision motion platform to emit light and control the light source, while improving the light source and environment during the operation of the equipment to avoid interference from external light sources. The precision motion platform performs precise displacement, and the precision motion platform also has a sensor for receiving displacement information and sending it to the controller. The precision motion platform is controlled by the controller and performs image acquisition after moving to a specified position. The controller is a servo controller, which connects the computer and the precision motion platform to realize the transmission of motion instructions and feedback of displacement information.

[0026] Specifically, the specific steps of obtaining the image set to be imaged obtained by sub-pixel sampling are: The imaging device receives the motion command and moves the precision motion platform to the specified position , the microscopic imaging unit collects images at the specified position ,Then, take the specified position as the reference point, and collect the adjacent positions of the reference point according to the template searched by the adjacent position Image At this time, the precision motion platform has moved to the adjacent position of the reference point , obtain the displacement information from the current position to other adjacent positions, collect images of each position through the precision motion platform, and form a sequence image set , this image set is the input data of the image reconstruction model.

[0027] Since the displacement of the precision motion platform is less than one pixel, the image collected after each displacement has sub-pixel differences in space. This embodiment obtains more spatial information by acquiring displacement information from the current position to other adjacent positions.

[0028] Step 2: Construct an image reconstruction model and train the image reconstruction model to obtain an optimized image reconstruction model.

[0029] The image reconstruction model includes a generator and a discriminator. The generator includes a variable sequence fusion module, a multi-dimensional attention mechanism module, and a residual graph feedback module. Figure 2 As shown in the figure, the sample image set is first subjected to variable sequence information fusion processing to learn the shallow information in the image and obtain a fused image sequence. Then, the deep features of the fused image sequence are extracted and mapped based on the multi-dimensional attention mechanism to obtain a deep feature sequence. Finally, the residual map is used to feedback and correct the image data. The discriminator performs discriminative feedback based on the generator results and adjusts the image reconstruction model parameters.

[0030] Step 2.1: Perform variable sequence information fusion processing on the image set.

[0031] Step 2.1.1: Screen the image set and classify them according to their relative positions in the pixel space to obtain image sequences of different groups.

[0032] In a specific implementation, since there are sub-pixel displacement differences between each group of images, the subsequent feature extraction method of each group of images is different, so it is necessary to perform group classification of images before feature extraction. Figure 3 As shown in the variable sequence fusion module, for the variable sequence microscopic images screened in the 5×5 search space, the path image sequence, i.e., the image set The classification is carried out according to their relative positions in the pixel space. The specific classification measures are as follows: the initial image, i.e. the image at the center of the pixel space, is used as the reference image as the first group , divided into the second group according to the difference in sub-pixel inter-frame displacement between the microscopic image and the reference image and the third group There is a 1 / 5 sub-pixel displacement difference between the images included in the second group and the reference image in the horizontal or vertical direction. According to the pixel space image position distribution, it can be seen that the number of microscopic images in this group is no more than 8. There is a 2 / 5 sub-pixel displacement difference between the images included in the third group and the reference image in the horizontal or vertical direction. According to the pixel space image position distribution, it can be seen that the number of microscopic images in this group is no more than 16, thereby realizing the fixed-group grading of variable sequence microscopic images.

[0033] Step 2.1.2: Perform shallow feature learning on different groups of image sequences respectively, and perform feature cascading on each group of shallow feature learning results to obtain a fused image sequence.

[0034] In a specific embodiment, Figure 3 As shown in the figure, group convolution processing is performed on different groups of image sequences: for the reference image of the first group, it is sent to the 3×3 convolution layer and the 1×1 convolution layer for shallow feature learning and processing, and for the second group and the third group Due to the uncertainty of the number of images, the channel dimension is first determined by pooling compression. Here, the maximum pooling and average pooling of the channel dimension are used for dimension compression to form 2-channel image compression feature data, which are respectively sent to the 3×3 convolution layer and the 1×1 convolution layer for shallow feature learning and processing. The features of the three groups of images after group convolution processing are further cascaded to form a 3-channel dimensional image feature map.

[0035] In the case of lack of image data in the second group, this module will replicate the features of the third group after the convolution operation, so as to ensure the feature dimension after feature cascading. Finally, multi-level feature extraction and dimensionality processing of variable sequence microscopic image data are realized, laying the foundation for subsequent deep feature extraction.

[0036] Step 2.2: After the variable sequence information is fused, the deep features of the image need to be extracted and mapped for subsequent reconstruction. When extracting and mapping the deep features, the residual dense block with the batch normalization layer removed is introduced into the multi-dimensional attention mechanism to solve the problem of loss of detail feature information and gradient disappearance or explosion.

[0037] In a specific implementation, this embodiment designs a multi-dimensional attention mechanism (MDAM) module, the specific structure is as follows: Figure 4 Input features , in view of the frequency and spatial attributes of different channel features, Fourier channel attention mechanism (FCA) and spatial attention mechanism (SA) are adopted. In FCA, different channels of feature maps capture the feature representation of input data at multiple levels and abstraction levels, especially emphasizing the importance of frequency domain features, which can reveal the structure and texture information of images at different levels. SA evaluates the importance of channel vectors of each spatial region in the input features and weights each spatial element, thereby enhancing the attention to key pixels, so that each region in the feature space obtains different attention weights according to its relevance to the target task. Finally, the network response to image features is finely adjusted from different dimensions. In the feature level dimension, the residual dense block (RDB) module is introduced to address the problems of loss of detail feature information, gradient disappearance or explosion, etc. The residual dense block is mainly divided into residual learning and dense connection. The cross-layer information transfer of the residual learning strategy alleviates the gradient disappearance problem in deep networks and helps the model better learn the identity mapping, thereby more accurately restoring the details of high-resolution images. Dense connections are implemented inside the block. Using the continuous memory mechanism, the input of each layer includes not only the output of the previous layer, but also the output of all previous layers, which greatly enhances the flow of information and ensures that the deepest network can also access rich local feature information. It is worth noting that in the field of image super-resolution, the BN layer may destroy the original contrast information of the image, which will have an adverse effect on the generation results. By removing the BN layer, the image details and contrast can be better preserved, thereby improving the resolution while ensuring the naturalness and authenticity of the image quality. Therefore, in order to further optimize the performance and retain the original contrast information of the image, this algorithm removes the batch normalization (BN) layer in the residual dense block, and the same measures are taken in other modules of the network (including other modules of the generator and the discriminator).

[0038] Step 2.3: Use the residual map to perform feedback correction on the image data.

[0039] In a specific implementation, the residual graph feedback process is as follows: Figure 5 As shown in the figure, in the residual feedback part, it includes downsampling feedback and upsampling fusion. The original high-resolution image is downsampled to obtain a downsampled image of the high-resolution image, which together with the reconstructed downsampled image constitutes the residual image. . The residual graph is obtained by upsampling operation , and after correction, the reconstructed image is obtained. The reconstructed downsampled image is obtained by downsampling the reconstructed image. In order to realize the feedback correction of the original low-resolution image data. The residual image feedback module structure is as follows Figure 6 As shown in Figure 2, this module can be divided into two parts: sub-pixel convolution and residual feedback. Input features Sub-pixel convolution is used to perform multi-view upsampling feedback operations. After the residual image is fed back, the corrected reconstructed image is obtained. Among them, in the sub-pixel convolution part, a sub-pixel convolution layer, two standard convolution layers and two activation functions are included to achieve effective upsampling of the image. It is divided into three steps. First, the feature extraction of the low-resolution image is performed, and then the convolution layer is used to generate a high-resolution feature map. Then the pixels of the feature map are rearranged, and finally a high-resolution image is obtained.

[0040] Step 2.4: The discriminator provides discriminative feedback based on the generator results.

[0041] Step 2.4.1: Use the convolutional feature extraction unit to gradually refine and amplify the feature information of the image.

[0042] In a specific implementation, the discriminator network constructed in this embodiment is composed of four main convolutional feature extraction units, which are intended to gradually refine and amplify the feature information of the image for internal processing in the discriminator to obtain the probability of authenticity, so as to accurately identify the image generated by the generator and the real high-resolution image as much as possible, and feed it back to the generator, thereby promoting the generator to produce higher quality super-resolution images. Each feature extraction unit is composed of two convolutional layers to achieve compression of the feature map dimension and increase the number of feature channels. At the end of the network, the features are further processed through two fully connected layers, and finally the authenticity probability of the image is output through the Sigmoid function, providing accurate feedback for the generator, promoting it to generate more refined and realistic super-resolution images.

[0043] Step 2.4.2: Calculate the loss function and generate discriminative feedback based on the calculation results.

[0044] Specifically, the loss function used in this embodiment combines three loss functions: content loss, adversarial loss, and frequency loss. It not only helps the model to achieve better performance on specific tasks, but also promotes stable training and optimization efficiency of the model.

[0045] Step 3: Use the optimized image reconstruction model to process the image set to obtain a reconstructed image.

[0046] Embodiment 2: Embodiment 2 of the present invention provides a microscopic imaging system based on deep learning, comprising: An image acquisition module is configured to acquire an image set to be imaged obtained by sub-pixel sampling; The reconstruction model training module is configured to construct an image reconstruction model and train the image reconstruction model to obtain an optimized image reconstruction model, wherein the image reconstruction model includes a generator and a discriminator. The generator first performs variable sequence information fusion processing on the sample image set, learns the shallow information in the image, and obtains a fused image sequence. Then, based on the multi-dimensional attention mechanism, the deep-level features of the fused image sequence are extracted and mapped to obtain a deep feature sequence. Finally, the residual map is used to feedback and correct the image data. The discriminator performs discrimination feedback based on the generator result and adjusts the image reconstruction model parameters. The image reconstruction module is configured to process the image set using the optimized image reconstruction model to obtain a reconstructed image.

[0047] Embodiment three: Embodiment 3 of the present invention provides a medium on which a program is stored. When the program is executed by a processor, the steps in the deep learning-based microscopic imaging method as described in Embodiment 1 of the present invention are implemented.

[0048] Embodiment 4: Embodiment 4 of the present invention provides a device, including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the deep learning-based microscopic imaging method as described in Embodiment 1 of the present invention are implemented.

[0049] The steps involved in the above embodiments 2, 3 and 4 correspond to the method embodiment 1. For the specific implementation methods, please refer to the relevant description part of embodiment 1.

[0050] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0051] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A microscopic imaging method based on deep learning, characterized in that: The following steps are involved: Acquire an image set to be imaged obtained by sub-pixel sampling; Construct an image reconstruction model and train it to obtain an optimized image reconstruction model, wherein the image reconstruction model includes a generator and a discriminator. The generator first performs variable sequence information fusion processing on the sample image set, learns the shallow information in the image, and obtains a fused image sequence. Then, based on the multi-dimensional attention mechanism, the deep-level features of the fused image sequence are extracted and mapped to obtain a deep feature sequence. Finally, the residual map is used to feedback and correct the image data. The discriminator performs discriminative feedback based on the generator results and adjusts the image reconstruction model parameters. The image set is processed using the optimized image reconstruction model to obtain a reconstructed image.

2. The deep learning-based microscopic imaging method according to claim 1, characterized in that: An image set is acquired through an imaging device, which includes a microscopic imaging unit, a composite light source unit, a precision motion platform and a controller. The microscopic imaging unit is used to capture image data, and is composed of a CMOS camera and a microscope. The composite light source units are installed on both sides of the precision motion platform to control the light source and improve the light source and environment during the operation of the equipment. The precision motion platform performs precise displacement. The precision motion platform is controlled by the controller and performs image acquisition after moving to the specified position. The controller connects the computer and the precision motion platform to realize the transmission of motion instructions and the feedback of information.

3. The deep learning-based microscopic imaging method according to claim 2, characterized in that: The specific steps of obtaining the image set to be imaged by sub-pixel sampling are: The imaging device receives the motion command and moves the precision motion platform to the specified position, collects the image of the specified position, and then uses the specified position as the reference point and collects the image of the adjacent position of the reference point according to the template of the adjacent position search. At this time, the precision motion platform has moved to the adjacent position of the reference point, and obtains the displacement information from the current position to other adjacent positions. The images of each position are collected by the precision motion platform to form a sequence image set.

4. The deep learning-based microscopic imaging method according to claim 1, characterized in that: The specific steps of performing variable sequence information fusion processing on the image set are: The image set is screened and classified according to the relative positions in the pixel space to obtain image sequences of different groups; Shallow feature learning is performed on different groups of image sequences, and each group of shallow feature learning results is concatenated to obtain a fused image sequence.

5. The deep learning-based microscopic imaging method according to claim 1, characterized in that: When extracting and mapping deep features, a residual dense block with batch normalization layer removed is introduced into the multi-dimensional attention mechanism to solve the problems of loss of detail feature information and gradient disappearance or explosion.

6. The deep learning-based microscopic imaging method according to claim 1, characterized in that: The specific steps of using the residual map to feedback correct the image data are: The image data is upsampled and fed back from multiple perspectives through sub-pixel convolution to obtain an upsampled image, which is then upsampled and fused after feedback from the residual map to obtain a corrected reconstructed image.

7. The deep learning-based microscopic imaging method according to claim 1, characterized in that: The specific steps for the discriminator to perform discriminative feedback based on the generator results are: The convolution feature extraction unit is used to gradually refine and amplify the feature information of the image; Calculate the loss function and generate discriminative feedback based on the calculation results.

8. A microscopic imaging system based on deep learning, characterized in that: include: An image acquisition module is configured to acquire an image set to be imaged obtained by sub-pixel sampling; The reconstruction model training module is configured to construct an image reconstruction model and train the image reconstruction model to obtain an optimized image reconstruction model, wherein the image reconstruction model includes a generator and a discriminator. The generator first performs variable sequence information fusion processing on the sample image set, learns the shallow information in the image, and obtains a fused image sequence. Then, based on the multi-dimensional attention mechanism, the deep-level features of the fused image sequence are extracted and mapped to obtain a deep feature sequence. Finally, the residual map is used to feedback and correct the image data. The discriminator performs discrimination feedback based on the generator result and adjusts the image reconstruction model parameters. The image reconstruction module is configured to process the image set using the optimized image reconstruction model to obtain a reconstructed image.

9. A computer-readable storage medium, characterized in that: A plurality of instructions are stored therein, and the instructions are suitable for being loaded by a processor of a terminal device and executing the deep learning-based microscopic imaging method described in any one of claims 1 to 7.

10. A terminal device, characterized in that: It includes a processor and a computer-readable storage medium, the processor is used to implement each instruction; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded by the processor and executing the deep learning-based microscopy imaging method described in any one of claims 1 to 7.

Images