A method for image processing of neural restoration in experimental animals

Through the experimental animal nerve recovery image super-resolution model, the multi-layer perceptron structure is used to achieve high-quality image reconstruction with 4 times resolution, which solves the problem of insufficient image resolution in the existing technology, and improves the accuracy of neuroscience research and Data quality for clinical applications.

CN118587095BActive Publication Date: 2024-11-22SHENYANG MEDICAL COLLEGE
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202411068856.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-06
Publication Date
2024-11-22
Estimated Expiration
2044-08-06

AI Technical Summary

Technical Problem

The prior art is difficult to effectively improve the resolution of neural recovery images in experimental animals, which limits the accuracy and data quality of neuroscience research, especially in the study of nerve damage repair and regeneration mechanisms.

Method used

A super-resolution model for neural recovery images of experimental animals is proposed. Through first-order and second-order initialization modules, noise reduction modules and high-resolution feature optimization modules, a multi-layer perceptron structure is used to achieve high-quality image reconstruction with 4 times resolution. Significantly reduces the computational complexity.

Benefits of technology

It realizes 4 times high-quality super-resolution image processing, improves the clarity and detail level of images, supports more accurate neuroscience research and clinical applications, and reduces computational complexity.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118587095B_ABST
    Figure CN118587095B_ABST
Patent Text Reader

Abstract

The present invention proposes a method for processing an experimental animal neural restoration image, and relates to the field of image super-resolution. The present invention proposes an experimental animal neural restoration image processing flow, including collecting experimental animal neural restoration images, constructing a first-order initialization module, constructing a first-order noise reduction module, constructing a second-order initialization module, constructing a second-order noise reduction module, constructing a high-resolution feature optimization module, constructing an experimental animal neural restoration image super-resolution model, training the experimental animal neural restoration image super-resolution model and processing the experimental animal neural restoration image; at the same time, an experimental animal neural restoration image super-resolution model is proposed, wherein the high-resolution feature optimization module realizes high-quality experimental animal neural restoration image reconstruction with 4 times resolution through an efficient multi-layer perceptron structure.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of image super-resolution, and in particular relates to a method for processing experimental animal nerve restoration images. Background Art

[0002] Super-resolution processing of experimental animal neural recovery images is of great significance for neuroscience research. First, super-resolution technology can significantly improve the clarity and detail level of images, allowing researchers to more accurately observe and analyze the microstructure of neurons and their connections. This is crucial for understanding the repair process after neural injury, the remodeling of neural circuits, and the mechanism of neuronal regeneration.

[0003] Secondly, by improving image resolution, researchers can more effectively identify and quantify subtle changes after nerve injury. This includes the regeneration of nerve fibers, synapse formation, and the response of glial cells, thus providing more reliable data support and quantitative analysis. This high-resolution image data helps to evaluate the effects of different treatments and promote the development and optimization of nerve repair therapies.

[0004] Finally, super-resolution images can be used to build more accurate neural network models to simulate the complex functions and behaviors of the nervous system. High-quality image data provides richer training data for machine learning and artificial intelligence technologies, thereby improving the accuracy and predictive ability of the models. This not only promotes basic neuroscience research, but also plays an important role in clinical applications such as neural regenerative medicine and the diagnosis and treatment of neurological diseases. Summary of the invention

[0005] The invention provides a method for processing an experimental animal neural restoration image, aiming to propose an experimental animal neural restoration image super-resolution model, wherein a first-order initialization module initializes a 2-fold resolution feature map, a first-order denoising module is used to improve the clarity and detail level of the 2-fold resolution feature map, and then a second-order initialization module is used to initialize a 4-fold resolution feature map, a second-order denoising module is used to improve the clarity and detail level of the 4-fold resolution feature map, and finally a high-resolution feature optimization module realizes high-quality image reconstruction with 4-fold resolution through an efficient multi-layer perceptron structure, while significantly reducing the computational complexity. The experimental animal neural restoration image super-resolution model is used for experimental animal neural restoration image processing, and a 4-fold high-quality super-resolution experimental animal neural restoration image can be obtained.

[0006] The present invention aims to propose a super-resolution model of experimental animal neural restoration images and provide an experimental animal neural restoration image processing method, comprising the following steps:

[0007] S1. Collection of experimental animal neural recovery images. During the experimental animal neural recovery cycle, regularly make stained tissue sections, and observe the stained tissue sections through an electron microscope to obtain multiple high-resolution experimental animal neural recovery images, and then perform four-fold downsampling to obtain multiple low-resolution experimental animal neural recovery images, and save them as multiple training pairs corresponding to the multiple high-resolution experimental animal neural recovery images, thereby forming a super-resolution dataset of experimental animal neural recovery images;

[0008] S2, build a first-order initialization module, including convolution and upsampling;

[0009] S3, construct a first-order denoising module, including a convolution block;

[0010] S4, build a second-order initialization module, including transposed convolution;

[0011] S5, construct a second-order denoising module, including a convolution block;

[0012] S6, constructing a high-resolution feature optimization module, including a cyclic feature block;

[0013] S7. Construct an experimental animal neural restoration image super-resolution model, including an input, a first-order initialization module, a first-order denoising module, a second-order initialization module, a second-order denoising module, a high-resolution feature optimization module and an output;

[0014] S8. Train the experimental animal neural restoration image super-resolution model, use the experimental animal neural restoration image super-resolution dataset to train the experimental animal neural restoration image super-resolution model, and divide the training set and the test set into seven to three parts;

[0015] S9, experimental animal neural recovery image processing, selecting low-resolution experimental animal neural recovery images that need super-resolution, inputting them into the experimental animal neural recovery image super-resolution model, and obtaining super-resolution experimental animal neural recovery images.

[0016] Preferably, in step S2, for the first-order initialization module, a low-resolution experimental animal neural recovery image is input , , h, w and 3 are low-resolution experimental animal neural recovery images The height, width and channels of Input to 3×3 convolution and ReLU function to get the feature map , , compared with low-resolution experimental animal neural recovery images , feature map The channel changes to , the feature map Input to 2 times upsampling to get feature map , , compared with the feature map , feature map The height and width of the feature map are increased to 2 times. It is the output feature map of the first-order initialization module.

[0017] Preferably, in step S3, for the first-order denoising module, the output of the first-order initialization module is first obtained. , , 2h, 2w and yes The height, width and channels of Noise characteristic diagram of , , and The dimensions are the same, Each position The value is , Representation interval The random number generated by uniform distribution in and Add element by element to get , , and then get the feature map , , It is the first convolution block, which consists of convolution operation, normalization operation and activation function. , and then get the feature map , , is the maximum pooling layer, , and then get the feature map , , is the second convolutional block, , and then get the feature map , , , and then get the feature map , , , It is the third convolution block, and then we get the feature map , , , is the first transposed convolution, and then and Add element by element to get the feature map , , and then get the feature map , , , is the second transposed convolution, and Add element by element to get the feature map , , and then get the feature map , , is the third transposed convolution, , and then get the feature map , , , It is the fourth transposed convolution, feature map It is the output feature map of the first-order denoising module.

[0018] Preferably, in step S4, for the second-order initialization module, a low-resolution experimental animal neural recovery image is input , , h, w and 3 are low-resolution experimental animal neural recovery images The height, width and channels of Input to the transposed convolution and ReLU function to get the feature map , , compared with low-resolution experimental animal neural recovery images , feature map The channel changes to , The height and width of the feature map are increased to 2 times. Input to 2 times upsampling to get feature map , , compared with the feature map , feature map The height and width of the feature map are increased to 2 times. It is the output feature map of the second-order initialization module.

[0019] Preferably, in step S5, for the second-order denoising module, firstly obtain the output feature map of the second-order initialization module And the output feature map of the first-order denoising module , the feature map and feature map Add element by element to get the feature map , , 4h, 4w and yes height, width and passages, It is the input feature map of the second-order denoising module, and the feature map is calculated. Noise characteristic diagram of , , and The dimensions are the same, Each position The value is , Representation interval The random number generated by uniform distribution in and Add element by element to get , , and then get the feature map , , It is the first convolution block, which consists of convolution operation, normalization operation and activation function. , and then get the feature map , , is the maximum pooling layer, , and then get the feature map , , is the second convolutional block, , and then get the feature map , , , and then get the feature map , , , It is the third convolution block, and then we get the feature map , , , is the first transposed convolution, and then and Add element by element to get the feature map , , and then get the feature map , , , is the second transposed convolution, and Add element by element to get the feature map , , and then get the feature map , , is the third transposed convolution, represents element-by-element addition, , feature map is the output of the second-order noise reduction module.

[0020] Preferably, in step S6, for the high-resolution feature optimization module, a cyclic feature block is first constructed. , loop feature block It consists of a normalization layer, a leaky ReLU activation function, a deformable convolution, a spatial attention layer, a random dropout layer, a normalization layer, a leaky ReLU activation function, a deformable convolution, and a cyclic feature block. It also contains a skip connection to connect the loop feature block Input and loop feature blocks The output of the last deformable convolution in is added element by element, and then the output feature map of the second-order denoising module is As the input feature map of the high-resolution feature optimization module, , 4h, 4w and yes The height, width and channel of , and then get the feature map , , , It is a 3×3 convolution, which transforms the feature map Input into multiple cascaded cyclic feature blocks to obtain feature maps , , and then generate the feature map The corresponding polar coordinate plot , , for the feature map Every position , using the radius and angle express, , , then the position Radius Calculated as ,Location Angle Calculated as , is the inverse tangent function of two variables, which forms a polar coordinate plot by calculating the radius and angle at all positions , and then the polar coordinate plot and feature map Add element by element to get the feature map , , and then obtain super-resolution experimental animal neural recovery images , , , It consists of a fully connected layer, a ReLU activation function, and layer normalization. Super-resolution experimental animal neural restoration image is the output of the high-resolution feature optimization module.

[0021] Preferably, in step S7, for the experimental animal neural recovery image super-resolution model, a low-resolution experimental animal neural recovery image is input , , h, w and 3 are low-resolution experimental animal neural recovery images The height, width and channels of low-resolution experimental animal neural recovery images Input to the first-order initialization module to obtain the feature map , , Represents the number of channels, and the feature map Input to the first-order denoising module to obtain the feature map , , restore low-resolution experimental animal neural images Input to the second-order initialization module to obtain the feature map , , the feature map and feature map Add element by element to get the feature map , the feature map Input to the second-order denoising module to obtain the feature map , , the feature map Input into the high-resolution feature optimization module to obtain super-resolution experimental animal neural recovery images , , compared with low-resolution experimental animal neural recovery images , Super-resolution neural recovery images of experimental animals The width and height are increased to 4 times.

[0022] Compared with the prior art, the present invention has the following technical effects:

[0023] The technical solution provided by the present invention proposes a super-resolution model for experimental animal neural restoration images, wherein a first-order initialization module initializes a 2-fold resolution feature map, and a first-order denoising module is used to improve the clarity and detail level of the 2-fold resolution feature map, and then a second-order initialization module is used to initialize a 4-fold resolution feature map, and a second-order denoising module is used to improve the clarity and detail level of the 4-fold resolution feature map, and finally a high-resolution feature optimization module realizes high-quality image reconstruction with 4 times the resolution through an efficient multi-layer perceptron structure, while significantly reducing the computational complexity. The super-resolution model for experimental animal neural restoration images is used for experimental animal neural restoration image processing, and 4-fold high-quality super-resolution experimental animal neural restoration images can be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a flowchart of the experimental animal nerve recovery image processing provided by the present invention.

[0025] Figure 2It is a structural diagram of the cyclic feature block CFBlock in the high-resolution feature optimization module provided by the present invention.

[0026] Figure 3 This is a structural diagram of the experimental animal neural restoration image super-resolution model provided by the present invention.

[0027] Figure 4 This is a super-resolution effect diagram of the experimental animal nerve recovery image provided by the present invention. DETAILED DESCRIPTION

[0028] The present invention aims to propose a method for processing experimental animal neural restoration images, and proposes an experimental animal neural restoration image super-resolution model, wherein a first-order initialization module initializes a 2-fold resolution feature map, and a first-order denoising module is used to improve the clarity and detail level of the 2-fold resolution feature map, and then a second-order initialization module is used to initialize a 4-fold resolution feature map, and a second-order denoising module is used to improve the clarity and detail level of the 4-fold resolution feature map, and finally a high-resolution feature optimization module realizes 4-fold resolution high-quality image reconstruction through an efficient multi-layer perceptron structure, while significantly reducing the computational complexity, and the experimental animal neural restoration image super-resolution model is used for experimental animal neural restoration image processing, so that a 4-fold high-quality super-resolution experimental animal neural restoration image can be obtained.

[0029] See also Figure 1 As shown, a method for processing images of experimental animal neural recovery in an embodiment of the present application is as follows:

[0030] S1. Collection of experimental animal nerve recovery images. Using mice as experimental animals, regularly prepare mouse brachial plexus nerve staining tissue sections during the recovery period of the mouse brachial plexus injury for 11 weeks, and observe the mouse brachial plexus nerve staining tissue sections under an electron microscope to obtain 220 high-resolution experimental animal nerve recovery images, with both height and width of 2048. Then, four-fold downsampling is performed to obtain multiple low-resolution experimental animal nerve recovery images, with both height and width of 512. These images are saved as 220 training pairs corresponding to the multiple high-resolution experimental animal nerve recovery images, thereby forming a super-resolution dataset of experimental animal nerve recovery images.

[0031] S2, build a first-order initialization module, including 3×3 convolution and 2x upsampling;

[0032] S3, build a first-order denoising module, which includes 3 convolution blocks;

[0033] S4, construct a second-order initialization module, including 1 transposed convolution;

[0034] S5, construct a second-order denoising module, which includes three convolution blocks;

[0035] S6, construct a high-resolution feature optimization module, including 5 cyclic feature blocks;

[0036] S7. Construct a super-resolution model for neural restoration images of experimental animals. Its structure is as follows: Figure 3 As shown, it includes input, 1 first-order initialization module, 1 first-order denoising module, 1 second-order initialization module, 1 second-order denoising module, 1 high-resolution feature optimization module and output;

[0037] S8. Train the experimental animal neural restoration image super-resolution model, using 200 training pairs in the experimental animal neural restoration image super-resolution dataset to train the experimental animal neural restoration image super-resolution model, and divide the training set and the test set into seven to three;

[0038] S9. Experimental animal neural recovery image processing: select low-resolution experimental animal neural recovery images from the remaining 20 training pairs as low-resolution experimental animal neural recovery images that need super-resolution, input them into the experimental animal neural recovery image super-resolution model, and obtain super-resolution experimental animal neural recovery images.

[0039] Furthermore, in step S2, for the first-order initialization module, a low-resolution neural restoration image of an experimental animal is input. , , 512, 512 and 3 are low-resolution experimental animal neural recovery images The height, width and channels of Input to 3×3 convolution and ReLU function to get the feature map , , compared with low-resolution experimental animal neural recovery images , feature map The channel changes to 64, and the feature map Input to 2 times upsampling to get feature map , , compared with the feature map , feature map The height and width of the feature map are increased to 2 times. It is the output feature map of the first-order initialization module.

[0040] Further, in step S3, for the first-order denoising module, the output of the first-order initialization module is first obtained. , , 1024, 1024 and 64 are The height, width and channels of Noise characteristic diagram of , , and The dimensions are the same, Each position The value is , Representation interval The random number generated by uniform distribution in and Add element by element to get , , and then get the feature map , , It is the first convolution block, which consists of convolution operation, normalization operation and activation function. , and then get the feature map , , is the maximum pooling layer, , and then get the feature map , , is the second convolutional block, , and then get the feature map , , , and then get the feature map , , , It is the third convolution block, and then we get the feature map , , , is the first transposed convolution, and then and Add element by element to get the feature map , , and then get the feature map , , , is the second transposed convolution, and Add element by element to get the feature map , , and then get the feature map , , is the third transposed convolution, , and then get the feature map , , , It is the fourth transposed convolution, feature map It is the output feature map of the first-order denoising module.

[0041] Further, in step S4, for the second-order initialization module, the low-resolution experimental animal neural recovery image is input , , 512, 512 and 3 are low-resolution experimental animal neural recovery images The height, width and channels of Input to the transposed convolution and ReLU function to get the feature map , , compared with low-resolution experimental animal neural recovery images , feature map The channel changes to 64, The height and width of the feature map are increased to 2 times. Input to 2 times upsampling to get feature map , , compared with the feature map , feature map The height and width of the feature map are increased to 2 times. It is the output feature map of the second-order initialization module.

[0042] Further, in step S5, for the second-order denoising module, the output feature map of the second-order initialization module is first obtained. And the output feature map of the first-order denoising module , the feature map and feature map Add element by element to get the feature map , , 2048, 2048 and 64 are height, width and passages, It is the input feature map of the second-order denoising module, and the feature map is calculated. Noise characteristic diagram of , , and The dimensions are the same, Each position The value is , Representation interval The random number generated by uniform distribution in and Add element by element to get , , and then get the feature map , , It is the first convolution block, which consists of convolution operation, normalization operation and activation function. , and then get the feature map , , is the maximum pooling layer, , and then get the feature map , , is the second convolutional block, , and then get the feature map , , , and then get the feature map , , , It is the third convolution block, and then we get the feature map , , , is the first transposed convolution, and then and Add element by element to get the feature map , , and then get the feature map , , , is the second transposed convolution, and Add element by element to get the feature map , , and then get the feature map , , is the third transposed convolution, represents element-by-element addition, , feature map is the output of the second-order noise reduction module.

[0043] Further, in step S6, for the high-resolution feature optimization module, a cyclic feature block is first constructed , whose structure is as follows Figure 2 As shown, the loop feature block It consists of a normalization layer, a leaky ReLU activation function, a deformable convolution, a spatial attention layer, a random dropout layer, a normalization layer, a leaky ReLU activation function, a deformable convolution, and a cyclic feature block. It also contains a skip connection to connect the loop feature block Input and loop feature blocks The output of the last deformable convolution in is added element by element, and then the output feature map of the second-order denoising module is As the input feature map of the high-resolution feature optimization module, , 2048, 2048 and 64 are The height, width and channel of , and then get the feature map , , , It is a 3×3 convolution, which transforms the feature map Input to 5 cascaded cyclic feature blocks to obtain the feature map , , and then generate the feature map The corresponding polar coordinate plot , , for the feature map Every position , using the radius and angle express, , , then the position Radius Calculated as ,Location Angle Calculated as , is the inverse tangent function of two variables, which forms a polar coordinate plot by calculating the radius and angle at all positions , and then the polar coordinate plot and feature map Add element by element to get the feature map , , and then obtain super-resolution experimental animal neural recovery images , , , It consists of a fully connected layer, a ReLU activation function, and layer normalization. Super-resolution experimental animal neural restoration image is the output of the high-resolution feature optimization module.

[0044] Further, in step S7, for the experimental animal neural restoration image super-resolution model, the low-resolution experimental animal neural restoration image is input , , 512, 512 and 3 are low-resolution experimental animal neural recovery images The height, width and channels of low-resolution experimental animal neural recovery images Input to the first-order initialization module to obtain the feature map , , 64 represents the number of channels, and the feature map Input to the first-order denoising module to obtain the feature map , , restore low-resolution experimental animal neural images Input to the second-order initialization module to obtain the feature map , , the feature map and feature map Add element by element to get the feature map , the feature map Input to the second-order denoising module to obtain the feature map , , the feature map Input into the high-resolution feature optimization module to obtain super-resolution experimental animal neural recovery images , , compared with low-resolution experimental animal neural recovery images , Super-resolution neural recovery images of experimental animals The width and height are increased to 4 times.

[0045] Furthermore, the experimental animal neural restoration image super-resolution model was written in Python using the Pytorch framework. During the training process, a 64GB NVIDIA RTX 4090 GPU was used, and the Adam optimizer with a learning rate of 0.00005 was used.

[0046] further, Figure 4 This is a super-resolution image of the restored nerves of experimental animals. The nerves of experimental animals here refer to the brachial plexus nerves of mice. Figure 4 The left is a low-resolution image of the neural recovery of experimental animals. Figure 4 On the right is a high-resolution image of the neural recovery of experimental animals. Compared with the low-resolution image of the neural recovery of experimental animals, the height and width of the high-resolution image of the neural recovery of experimental animals are increased by four times.

[0047] The above are only preferred embodiments of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for processing images of experimental animal neural restoration, characterized in that: The following steps are involved: S1. Collection of experimental animal neural recovery images. During the experimental animal neural recovery cycle, regularly make stained tissue sections, and observe the stained tissue sections through an electron microscope to obtain multiple high-resolution experimental animal neural recovery images, and then perform four-fold downsampling to obtain multiple low-resolution experimental animal neural recovery images, and save them as multiple training pairs corresponding to the multiple high-resolution experimental animal neural recovery images, thereby forming a super-resolution dataset of experimental animal neural recovery images; S2. Construct a first-order initialization module, including convolution and upsampling; specifically: input low-resolution experimental animal neural recovery image , , h, w and 3 are low-resolution experimental animal neural recovery images The height, width and channels of Input to 3×3 convolution and ReLU function to get the feature map , , compared with low-resolution experimental animal neural recovery images , feature map The channel changes to , the feature map Input to 2 times upsampling to get feature map , , compared with the feature map , feature map The height and width of the feature map are increased to 2 times. is the output feature map of the first-order initialization module; S3. Construct a first-order denoising module, including a convolution block. Specifically, first obtain the output of the first-order initialization module. , , 2h, 2w and yes The height, width and channels of Noise characteristic diagram of , , and The dimensions are the same, Each position The value is , Representation interval The random number generated by uniform distribution in and Add element by element to get , , and then get the feature map , , It is the first convolution block, which consists of convolution operation, normalization operation and activation function. , and then get the feature map , , is the maximum pooling layer, , and then get the feature map , , is the second convolutional block, , and then get the feature map , , , and then get the feature map , , , It is the third convolution block, and then we get the feature map , , , is the first transposed convolution, and then and Add element by element to get the feature map , , and then get the feature map , , , is the second transposed convolution, and Add element by element to get the feature map , , and then get the feature map , , is the third transposed convolution, , and then get the feature map , , , It is the fourth transposed convolution, feature map It is the output feature map of the first-order denoising module; S4. Construct a second-order initialization module, including transposed convolution; specifically: input low-resolution experimental animal neural recovery image , , h, w and 3 are low-resolution experimental animal neural recovery images The height, width and channels of Input to the transposed convolution and ReLU function to get the feature map , , compared with low-resolution experimental animal neural recovery images , feature map The channel changes to , The height and width of the feature map are increased to 2 times. Input to 2 times upsampling to get feature map , , compared with the feature map , feature map The height and width of the feature map are increased to 2 times. is the output feature map of the second-order initialization module; S5. Construct a second-order denoising module, including a convolution block; specifically, first obtain the output feature map of the second-order initialization module And the output feature map of the first-order denoising module , the feature map and feature map Add element by element to get the feature map , , 4h, 4w and yes height, width and passages, It is the input feature map of the second-order denoising module, and the feature map is calculated. Noise characteristic diagram of , , and The dimensions are the same, Each position The value is , Representation interval The random number generated by uniform distribution in and Add element by element to get , , and then get the feature map , , It is the first convolution block, which consists of convolution operation, normalization operation and activation function. , and then get the feature map , , is the maximum pooling layer, , and then get the feature map , , is the second convolutional block, , and then get the feature map , , , and then get the feature map , , , It is the third convolution block, and then we get the feature map , , , is the first transposed convolution, and then and Add element by element to get the feature map , , and then get the feature map , , , is the second transposed convolution, and Add element by element to get the feature map , , and then get the feature map , , is the third transposed convolution, represents element-by-element addition, , feature map is the output of the second-order noise reduction module; S6, construct a high-resolution feature optimization module, including a cyclic feature block; specifically: for the high-resolution feature optimization module, first construct a cyclic feature block , loop feature block It consists of a normalization layer, a leaky ReLU activation function, a deformable convolution, a spatial attention layer, a random dropout layer, a normalization layer, a leaky ReLU activation function, a deformable convolution, and a cyclic feature block. It also contains a skip connection to connect the loop feature block Input and loop feature blocks The output of the last deformable convolution in is added element by element, and then the output feature map of the second-order denoising module is As the input feature map of the high-resolution feature optimization module, , 4h, 4w and yes The height, width and channel of , and then get the feature map , , , It is a 3×3 convolution, which transforms the feature map Input into multiple cascaded cyclic feature blocks to obtain feature maps , , and then generate the feature map The corresponding polar coordinate plot , , for the feature map Every position , using the radius and angle express, , , then the position Radius Calculated as ,Location Angle Calculated as , is the inverse tangent function of two variables, which forms a polar coordinate plot by calculating the radius and angle at all positions , and then the polar coordinate plot and feature map Add element by element to get the feature map , , and then obtain super-resolution experimental animal neural recovery images , , , It consists of a fully connected layer, a ReLU activation function, and layer normalization. Super-resolution experimental animal neural restoration image is the output of the high-resolution feature optimization module; S7. Construct a super-resolution model of experimental animal neural restoration images, including input, first-order initialization module, first-order denoising module, second-order initialization module, second-order denoising module, high-resolution feature optimization module and output; specifically: for the experimental animal neural restoration image super-resolution model, input low-resolution experimental animal neural restoration image , , h, w and 3 are low-resolution experimental animal neural recovery images The height, width and channels of low-resolution experimental animal neural recovery images Input to the first-order initialization module to obtain the feature map , , Represents the number of channels, and the feature map Input to the first-order denoising module to obtain the feature map , , restore low-resolution experimental animal neural images Input to the second-order initialization module to obtain the feature map , , the feature map and feature map Add element by element to get the feature map , the feature map Input to the second-order denoising module to obtain the feature map , , the feature map Input into the high-resolution feature optimization module to obtain super-resolution experimental animal neural recovery images , , compared with low-resolution experimental animal neural recovery images , Super-resolution neural recovery images of experimental animals The width and height are increased to 4 times; S8. Training an experimental animal neural restoration image super-resolution model, using an experimental animal neural restoration image super-resolution dataset to train an experimental animal neural restoration image super-resolution model, and dividing the training set and the test set into a ratio of seven to three; S9, experimental animal neural recovery image processing, selecting low-resolution experimental animal neural recovery images that need super-resolution, and inputting them into the experimental animal neural recovery image super-resolution model to obtain super-resolution experimental animal neural recovery images.

Citation Information

Patent Citations

  • Multi-stage progressive image denoising algorithm

    CN113808032A

  • Two-stage deep network image super-resolution reconstruction method suitable for multiple blurring kernels

    CN116188272A

  • Image super-resolution reconstruction method based on generative adversarial network

    CN118396853A

  • Tumor radiotherapy effect CT image processing method

    CN118397130A

  • Image super-resolution method and system

    WO2020238558A1