DW image super-resolution reconstruction method, system, computer equipment and medium

Abstract

The present invention provides a DW image super-resolution reconstruction method, system, computer device, and medium, belonging to the field of image processing technology. The method comprises: in Q space, calculating the similarity between the target direction and the remaining directions using cosine similarity, selecting DW image slices in N directions with the largest cosine similarity to form N adjacent images; utilizing an RDN network architecture to extract features of the target direction and the N adjacent direction images respectively; orthogonally projecting and fusing the target direction features with the features of the N adjacent direction images to generate fused features; combining the fused features with the high-resolution image coordinates using an implicit neural network module, mapping them into a continuous function, and employing an image-to-image learning approach, incorporating data fidelity loss, spatial gradient change loss, and structural similarity loss to reconstruct a super-resolution DW image of any scale. The present invention utilizes the above-mentioned method to improve the super-resolution quality of DW images.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a DW image super-resolution reconstruction method, system, computer equipment and medium. Background Art

[0002] Diffusion-weighted (DW) imaging utilizes the Brownian motion of water molecules for tissue imaging. Currently, it is the only technique that can noninvasively monitor water molecule diffusion within living tissues. Water molecules in living tissues are restricted in diffusion by nerves, cells, and surrounding tissue structures, and the distribution of their diffusion displacement can reflect the tissue's microstructure. Currently, DW imaging has been widely used to examine complex fibrous structures such as the brain, myocardium, and liver. The resolution of DW images is often limited by various factors. Firstly, obtaining high-resolution DW images requires more sophisticated imaging equipment, which places significant financial pressure on patients and hospitals. Secondly, from an imaging technology perspective, ensuring the accuracy of IVIM parameter mapping and reconstructing other dMRI modalities such as DTI typically requires the acquisition of diffusion-weighted images with multiple b-values ​​and motion directions, which inevitably increases acquisition time. However, long acquisition times can introduce numerous issues, such as motion artifacts. Consequently, image resolution is often reduced to shorten acquisition time. Many research methods have been proposed to obtain high-resolution DW images, such as EDSR and RDN. However, current DW image super-resolution reconstruction methods often only target integer or fixed multiples. Each reconstruction multiple requires retraining and saving the model, resulting in a waste of computing and storage resources, which greatly limits clinical applications.

[0003] In summary, using the same model to achieve super-resolution reconstruction of DW images at any spatial scale can meet the needs of doctors to observe images at different resolutions. At the same time, reconstructing high-resolution images with richer texture details can improve the doctor's diagnostic accuracy and be more in line with practical clinical applications. Summary of the Invention

[0004] The present invention aims to provide a method, system, computer device, and medium for super-resolution reconstruction of DW images. This method considers that the spatial structures of multi-directional DW image slices have certain similarities due to being located at the same location in the brain, while the grayscale distribution has certain differences due to the influence of gradient fields in different directions. Multi-directional projection fusion is used to enhance feature independence and address the problem of insufficient INR representation of detailed texture features. Furthermore, the INR point-to-point learning method is modified to an image-to-image learning method. Spatial gradient loss and structural similarity loss are added to the data fidelity loss to constrain the model to learn the spatial and structural information of the image, thereby reconstructing high-quality, high-resolution images with rich texture information.

[0005] To achieve the above object, the present invention provides a DW image super-resolution reconstruction method, comprising the following steps:

[0006] Step S1: In Q space, calculate the similarity between the target direction and the other directions by cosine similarity, and select the direction with the largest cosine similarity. DW image slices in different directions adjacent images;

[0007] Step S2: using the RDN network architecture to extract features of the target direction and N adjacent direction images respectively;

[0008] Step S3: performing orthogonal projection fusion on the target direction feature and the features of N adjacent direction images to generate a fusion feature;

[0009] Step S4: Combine the fusion features with the high-resolution image coordinates through an implicit neural network module, and map the image features and coordinates to image pixel values ​​through a multi-layer perceptron. In this way, the discrete image is represented by a continuous function through the multi-layer perceptron. An image-to-image learning method is adopted, and data fidelity loss, spatial gradient change loss, and structural similarity loss are added to reconstruct a super-resolution DW image of any scale.

[0010] In the implicit neural network module, the fusion features are obtained through a convolution layer to obtain the frequency and amplitude of the features, and the frequency information is upsampled and compared with the two-dimensional coordinates of the input high-resolution image. x and y After multiplication, the elements are added one by one, and the scale information of the grid size of the high-resolution image is mapped to the same dimension as the frequency information through a linear layer and then added as the phase information of the feature. Then, the sin and cos activation functions are used to obtain the periodic feature information. Finally, after upsampling with the amplitude information, it is multiplied and passed through a multi-layer perceptron (continuous function representor) to obtain the reconstructed high-resolution image.

[0011] Preferably, in step S1, the cosine similarity calculation formula is as follows:

[0012] ;

[0013] in, represents the cosine similarity, Represents the three-dimensional coordinate vector of the target direction, Represents the coordinate vectors of the remaining directions, express and The angle between represents the index of a direction in Q space, Represents the number of all directions in Q space.

[0014] Preferably, in step S2, the RDN network architecture includes extracting shallow features through two convolutions, extracting deep features through 16 RDB blocks, and then The convolution channel reduces the dimension, and finally the deep features and shallow features are added through the residual connection to obtain the features of the image.

[0015] Preferably, in step S3, the target direction feature is fused with the features of N adjacent direction images by orthogonal projection to generate a fused feature. The specific operation is:

[0016] Target direction feature To each adjacent direction feature Perform orthogonal projection to obtain the projection residual feature :

[0017] ;

[0018] All projected residual features Target direction features Splicing to generate fusion features:

[0019] ;

[0020] in, Represents the fusion feature, and N represents the number of adjacent images selected in the Q space.

[0021] Preferably, the resolution of the high-resolution image is 145x174.

[0022] Preferably, data fidelity loss as follows:

[0023] ;

[0024] in, Indicates the absolute value, represents the network super-resolution reconstructed image, represents the true label image;

[0025] Structural similarity loss as follows:

[0026] ;

[0027] ;

[0028] in, and Represents images respectively and images The average value of and They represent the variance of the two, represents the covariance between the two, and is a constant less than 0.003; Indicates the structural similarity between image Pred and image GT, with a value ranging from 0 to 1. Represents the structural similarity loss between image Pred and image GT;

[0029] Spatial gradient loss as follows:

[0030] ;

[0031] ;

[0032] ;

[0033] ;

[0034] ;

[0035] in, , Represents the Sobel operator in the horizontal and vertical directions respectively, Represents a splicing operation, For the predicted image Pred and Sobey operator and The splicing result of sliding multiplication represents the spatial gradient change of the predicted image Pred. For the real image GT and Sobey operator and The splicing result of sliding multiplication represents the real image GT spatial gradient change,

[0036] The total losses are as follows:

[0037] ;

[0038] in, represents the total loss, 、 、 represents a hyperparameter.

[0039] The present invention also provides a DW image super-resolution reconstruction system, comprising:

[0040] The multi-directional projection fusion module is used to calculate the similarity between the target direction and other directions through cosine similarity in Q space, select the DW image slices in the N directions with the largest cosine similarity to form a neighboring image, and use the RDN network architecture to extract the features of the target direction and the N neighboring direction images respectively; the target direction feature and the features of the N neighboring direction images are orthogonally projected and fused to generate a fused feature;

[0041] An implicit neural network module is used to combine the fusion features with the high-resolution image coordinates, map them into a continuous function, and use an image-to-image learning method to add data fidelity loss, spatial gradient change loss, and structural similarity loss to reconstruct super-resolution DW images of any scale;

[0042] The loss calculation module is used to calculate data fidelity loss, spatial gradient change loss, and structural similarity loss, and optimize the model based on these losses.

[0043] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory is used to store instructions, and the processor is used to execute the instructions to implement the above-mentioned DW image super-resolution reconstruction method.

[0044] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the above-mentioned DW image super-resolution reconstruction method when executed by a processor.

[0045] Therefore, the present invention adopts the above-mentioned DW image super-resolution reconstruction method, system, computer device and medium, and the beneficial technical effects are as follows:

[0046] (1) The method proposed in this paper, with the help of the multi-directional characteristics of DW, fully exploits the fact that the spatial structures of slices with different diffusion gradient directions are similar and the grayscale distribution is different, extracts the adjacent direction features and fuses them with the target direction features through projection, thereby enhancing the independence of the target direction feature information, thereby effectively overcoming the problem of lack of detailed texture feature representation due to the poor independence of INR features.

[0047] (2) Through implicit neural networks, a single model is used to perform super-resolution reconstruction of DW images at any spatial scale. At the same time, a new loss constraint model and a new image-to-image learning method are proposed, which enable the model to have the ability to perceive spatial changes and reconstruct detailed structural texture information, thereby improving the quality of reconstructed images. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 The image and grayscale distribution of the target direction and three different directions;

[0049] Figure 2 Flowchart of the DW image super-resolution reconstruction method;

[0050] Figure 3 is an implicit neural network module;

[0051] Figure 4 This is a diagram of the learning method of implicit neural network;

[0052] Figure 5 This is the result of 2.5x super-resolution reconstruction;

[0053] Figure 6 This is the result of 3.5x super-resolution reconstruction;

[0054] Figure 7 The feasibility of the projection method;

[0055] Figure 8 Visualize feature independence;

[0056] Figure 9 To verify the effectiveness of the loss function and learning method. DETAILED DESCRIPTION

[0057] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0058] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0059] Example 1

[0060] DW image is a 4D image that has both spatial and directional information. In order to use multi-directional information to assist in the super-resolution reconstruction of the spatial structure of the target direction in the INR framework, the histogram of the grayscale distribution of the Q-space target direction and the slices in three different directions is visualized and statistics are given as follows: Figure 1 .Depend on Figure 1It can be concluded that the spatial structure of multi-directional slices of DW images has certain similarities due to the same location in the brain, while the grayscale distribution has certain differences due to the influence of gradients in different directions. Therefore, the method proposed in this paper utilizes the characteristics of multi-directional slices with similar spatial structure and different grayscale distribution to project and fuse multi-directional information to solve the problem of insufficient representation of detailed texture features caused by poor feature independence in implicit neural representation. In addition, a new image-to-image learning method and a new loss function are used to constrain the INR model to learn the spatial variation and texture information of the DW image, thereby reconstructing high-quality high-resolution images.

[0061] like Figure 2 As shown in the figure, it is a flow chart of the DW image super-resolution reconstruction method. In the Q space, the similarity between the target direction and the other directions is calculated by cosine similarity, and the direction with the largest cosine similarity is selected. The DW image slices in each direction form N adjacent images, and the cosine similarity calculation formula is as follows:

[0062] ;

[0063] in, represents the cosine similarity, Represents the three-dimensional coordinate vector of the target direction, Represents the coordinate vectors of the remaining directions, express and The angle between represents the index of a direction in Q space, Represents the number of all directions in Q space.

[0064] Step S2: using the RDN network architecture to extract the features of the target direction and N direction images respectively;

[0065] For each input image, shallow features are first extracted through two convolutions , shallow features are extracted through 16 RDB blocks (residual dense blocks) to extract deep features, through Convolution channel dimensionality reduction, and then add deep features and shallow features through residual connection to obtain image features Each RDB consists of 8 Convolution and The activation function is formed by combining dense connections and residual connections. The first RDB Features of layer convolution Need to combine The features of each convolution layer in the previous RDB, that is, arrive Features, among which For the RDB output features. The first RDB output needs to be The convolution output of each layer in the RDB is concatenated in the channel dimension. After the convolution output is connected with the residual Add up to get the output features . Where [ ] represents the splicing operation, represents the activation function, Represents an RDB Convolution. Similarly, we can get the final image features of RDN. Indicates total D RDB blocks.

[0066] ;

[0067] ;

[0068] ;

[0069] Step S3: Perform orthogonal projection fusion on the target direction feature and N adjacent direction image features to generate a fusion feature.

[0070] Target direction feature To each adjacent direction feature Perform orthogonal projection to obtain the projection residual feature :

[0071] ;

[0072] All projected residual features Target direction features Splicing to generate fusion features:

[0073] ;

[0074] in, Represents the fusion feature, and N represents the number of selected adjacent directions.

[0075] Step S4: Combine the fusion features with the high-resolution image coordinates through the implicit neural network module, map the image features and coordinates to image pixel values ​​through the multi-layer perceptron, obtain a continuous function representation of the image, and adopt an image-to-image learning method, add spatial gradient change loss and structural similarity loss, and reconstruct a super-resolution DW image of any scale.

[0076] The implicit neural network module will be enriched by the feature projection fusion module feature representation Combined with HR coordinates, features and coordinates are mapped to pixel values. This process can represent discrete 2D images as continuous functions as follows:

[0077] ;

[0078] in, is the pixel value, Characterized by are coordinates, is the network parameter.

[0079] The specific network architecture is as follows Figure 3 , through a convolutional layer to extract the frequency and amplitude of the features. After upsampling the frequency information, it is compared with the two-dimensional coordinates of the input original high-resolution image. and The components are multiplied separately and added element by element. This coordinate addition method can obtain a better coordinate feature representation. Then, the scale information of the grid size of the original high-resolution image is mapped to the same dimension as the frequency information through a linear layer (Linear), and it is added as the phase information (Phase) of the feature. Next, the sin and cos activation functions are used to obtain the periodic change features that can characterize the high-frequency information. Finally, these features are multiplied with the upsampled amplitude information and output through a four-layer multi-layer perceptron (MLP) to obtain the reconstructed high-resolution image ( Pred ).

[0080] The sampling strategy used by traditional INR in learning image representation has limitations. Figure 4 As shown on the left, by sampling pixel values ​​in the high-resolution (HR) space as a supervisory signal, the model is constrained to learn the mapping relationship between features and coordinates. However, this point-to-point learning method makes it difficult for the model to capture the texture information and spatial gradient changes of the image, resulting in insufficient details and overly smooth reconstructed images. To solve this problem, the present invention proposes an image-to-image learning method, such as Figure 4 As shown on the right, the network generates high-resolution images through learning. Spatial gradient loss and structural similarity loss are introduced to constrain the model learning process, allowing the model to learn the spatial variation information and detailed texture information of the image. Through this new loss function and learning method, the model can reconstruct more realistic and detailed high-resolution images. The new loss function is as follows:

[0081] ;

[0082] Loss of data fidelity as follows:

[0083] ;

[0084] in, Indicates the absolute value, represents the network super-resolution reconstructed image, represents the true label image;

[0085] Structural similarity loss as follows:

[0086] ;

[0087] ;

[0088] in, and Represents images respectively and images The average value of and They represent the variance of the two, represents the covariance between the two, and is a constant less than 0.003; Indicates the structural similarity between image Pred and image GT, with a value ranging from 0 to 1. Represents the structural similarity loss between image Pred and image GT;

[0089] Spatial gradient loss as follows:

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] ;

[0095] in, , Represents the Sobel operator in the horizontal and vertical directions respectively, Represents a splicing operation, For the predicted image Pred and Sobey operator and The splicing result of sliding multiplication represents the spatial gradient change of the predicted image Pred. For the real image GT and Sobey operator and The splicing result of sliding multiplication represents the spatial gradient change of the real image GT.

[0096] The present invention will be further described below through specific examples.

[0097] Experiments were conducted on the HCP dataset. The original image was downsampled at multiple scales and then super-resolved back to the original resolution. Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) metrics were calculated, demonstrating that the proposed method achieved the best reconstruction results compared to seven comparison methods. The stability of the PSNR was also verified by comparing the Var (Var) of the PSNR on test data. The SSIM calculation method is identical to the aforementioned SSIM, differing in that the loss is computed using a square Gaussian kernel with hard-coded parameters and explicit data range input, outputting a tensor format compatible with automatic differentiation. Image quality was assessed using a circular Gaussian kernel and dynamic range estimation, with a direct scalar result return. Furthermore, visualizations were performed on the proposed method at 2.5x and 3.5x downsampling resolutions (58x69) and (41x49) back to the original resolution (145x174). The proposed method achieved the best reconstruction results compared to the seven comparison methods.

[0098] The calculation formulas for PSNR and Var are as follows:

[0099] ;

[0100] ;

[0101] in Represent the network reconstructed image and the real image respectively, H and W represent the height and width of the image, 、 Indicates that the reconstructed image and the real image are in the mth row and the The pixel value of the column. Represents the maximum pixel value in the image Pred, is the mean square error between image Pred and image GT. Var represents the variance of PSNR, represents the PSNR of the lth image in the test set, represents the mean PSNR of all test images, The smaller it is, the better the data stability.

[0102] Table 1 shows the quantitative index results of reconstruction of different methods. All results are b values. The super-resolution reconstruction results under . In-Scale in the table indicates the super-resolution scale contained in the training model, and Out-of-Scale indicates the super-resolution scale beyond the training range. In addition, the data set of the present invention is the HCP data set after denoising by patch2self. It can be seen from the table that the PSNR and SSIM of the method of the present invention are better than those of the comparison method at any scale, and the variance does not increase, and the data stability is good. In addition, as the super-resolution scale increases, the PSNR and SSIM of the present invention are more improved than those of the best comparison method LTE, and the method of the present invention has a better reconstruction effect on larger scales.

[0103] Table 1 PSNR and SSIM of reconstruction results of each method

[0104] ;

[0105] Figure 5 、 Figure 6 The images are 2.5x and 3.5x super-resolution reconstruction results respectively. The first row of the images contains the real HR images (GT) and the images reconstructed by the method of the present invention and other 7 comparison methods, namely ARSSR, OPESR, CLIT, LIIF, ITSRN, GAUSR, and LTE. The second and fourth rows of the images are the red and blue magnified areas in the first row respectively. The first columns of the third and fifth rows are the downsampled low-resolution images. The remaining columns are the residual images of the red and blue magnified areas and the real label magnified areas. The lighter the color of the residual image, the better the reconstruction effect. Figure 5 From the results of the 2.5 super-resolution reconstruction, it can be seen that the method of the present invention is better than other comparison methods in the detail texture area, and the residual map does not show structural texture, which proves that the reconstruction result is closer to the real GT image. Figure 6 The 3.5x super-resolution reconstruction results show that the proposed method significantly outperforms other methods in detail texture reconstruction, with the texture edge structure closer to the true (GT) image. The residual image also shows that the reconstruction of texture areas is significantly better than that of other comparison methods. Comparing the 2.5x and 3.5x super-resolution reconstruction results, it can be seen that the proposed method achieves better reconstruction results at large scales. Specifically, when more image information is lacking, it can reconstruct images that are closer to the true image, reconstructing more realistic and detailed textures.

[0106] In order to verify the feasibility of the projection fusion method of the present invention, Figure 7 The second feature A is projected onto the first feature B to obtain the projected feature of the third image. The second image is subtracted from the third image to obtain the projection residual feature map. It can be concluded that projection subtraction can indeed enhance feature independence and remove the influence of irrelevant features. Feature independence can be visualized as follows: Figure 8The present invention selects any point in the image after 2.0 times and 3.5 times downsampling, and calculates the feature similarity between the selected point and the eight surrounding points, where the first picture is the position of the selected point, the second picture is the cosine similarity between the selected point and the eight surrounding points in the original feature without projection fusion, and the third picture is the cosine similarity between the selected point and the eight surrounding points after projection fusion. From the comparison of the second and third pictures, it can be seen that the feature similarity between the selected point and the surrounding points is reduced, thereby enhancing the feature independence of the point. Enhanced feature independence can solve the problem of insufficient feature representation of texture information caused by poor INR feature independence.

[0107] In order to verify the effectiveness of the method of the present invention by adding loss and image-to-image learning, the present invention adds a new learning method and a new loss reconstruction result on the basis of the baseline model. Figure 9 The first row shows the ground-truth image and the reconstructed high-resolution image using the new learning method and loss. The second and third rows show the red magnified area and residual image of the red magnified area in the first row, respectively. The fourth row shows the residual image of the entire image. The blue number above represents the residual mean error; smaller is better. The results show that the proposed loss and learning method achieve the best results in the detailed texture area, with the highest PSNR and SSIM indicators and the lowest overall mean error. This proves that the proposed loss and image-to-image learning method using a desampling strategy can solve the problem that the original method cannot learn image texture and spatial information, thereby enabling the model to perceive image texture and spatial information.

[0108] Example 2

[0109] A DW image super-resolution reconstruction system, comprising:

[0110] The multi-directional projection fusion module is used to calculate the similarity between the target direction and other directions through cosine similarity in Q space, select the DW image slices in the N directions with the largest cosine similarity to form a neighboring image, and use the RDN network architecture to extract the features of the target direction and the N neighboring direction images respectively; the target direction features and the features of the N neighboring direction images are orthogonally projected and fused to generate fused features;

[0111] An implicit neural network module is used to combine the fusion features with the high-resolution image coordinates, map them into a continuous function, and use an image-to-image learning method to add data fidelity loss, spatial gradient change loss, and structural similarity loss to reconstruct super-resolution DW images of any scale;

[0112] The loss calculation module is used to calculate data fidelity loss, spatial gradient change loss, and structural similarity loss, and optimize the model based on these losses.

[0113] If the above functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0114] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0115] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, and then editing, interpreting, or processing in another suitable manner as necessary, and then storing it in a computer memory.

[0116] It is worth noting that the contents not elaborated in detail in the present invention are all prior art and are well known to those skilled in the art.

[0117] Therefore, the present invention adopts the above-mentioned DW image super-resolution reconstruction method, system, computer device and medium. This method considers that the spatial structure between multi-directional slices of the DW image has certain similarities due to the same location in the brain, while the grayscale distribution has certain differences due to the influence of gradient fields in different directions. Multi-directional projection fusion is used to enhance feature independence to solve the problem of insufficient INR ability to represent detailed texture features. At the same time, the INR point-to-point modification is converted to an image-to-image learning method and the spatial gradient loss and structural similarity loss constraint model are used to learn the spatial and structural information of the image, thereby reconstructing high-quality high-resolution images with rich texture information.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A DW image super-resolution reconstruction method, characterized in that: The following steps are involved: Step S1: In Q space, calculate the similarity between the target direction and the other directions by cosine similarity, and select the direction with the largest cosine similarity. DW image slices in different directions adjacent images; Step S2: using the RDN network architecture to extract features of the target direction and N adjacent direction images respectively; Step S3: performing orthogonal projection fusion on the target direction feature and the features of N adjacent direction images to generate a fusion feature; Step S4: Combining the fusion features with the high-resolution image coordinates through an implicit neural network module, mapping the image features and coordinates to image pixel values ​​through a multi-layer perceptron, obtaining a continuous function representation of the image, and adopting an image-to-image learning method, adding data fidelity loss, spatial gradient change loss, and structural similarity loss to reconstruct a super-resolution DW image of any scale; In the implicit neural network module, the fusion features are obtained through a convolution layer to obtain the frequency and amplitude of the features, and the frequency information is upsampled and compared with the two-dimensional coordinates of the input high-resolution image. x and y After multiplication, the elements are added one by one, and then the scale information of the grid size of the high-resolution image is mapped to the same dimension as the frequency information through a linear layer and added as the phase information of the feature. Then, the sin and cos activation functions are used to obtain the periodic feature information. Finally, after upsampling with the amplitude information, it is multiplied and passed through a multi-layer perceptron to obtain the reconstructed high-resolution image.

2. The DW image super-resolution reconstruction method according to claim 1, characterized in that: In step S1, the cosine similarity calculation formula is as follows: ； in, represents the cosine similarity, Represents the three-dimensional coordinate vector of the target direction, Represents the coordinate vectors of the remaining directions, express and The angle between represents the index of a direction in Q space, Represents the number of all directions in Q space.

3. The DW image super-resolution reconstruction method according to claim 1, characterized in that: In step S2, the RDN network architecture includes extracting shallow features through two convolutions, extracting deep features through 16 RDB blocks, and then The convolution channel reduces the dimension, and finally the deep features and shallow features are added through the residual connection to obtain the features of the image.

4. The DW image super-resolution reconstruction method according to claim 1, characterized in that: In step S3, the target direction feature is fused with the features of N adjacent direction images by orthogonal projection to generate a fused feature. The specific operation is as follows: Target direction feature To each adjacent direction feature Perform orthogonal projection to obtain the projection residual feature : ； All projected residual features Target direction features Splicing to generate fusion features: ； in, Represents the fusion feature, and N represents the number of adjacent images selected in the Q space.

5. The DW image super-resolution reconstruction method according to claim 1, characterized in that: The resolution of the high-resolution images is 145x174.

6. The DW image super-resolution reconstruction method according to claim 1, characterized in that: Loss of data fidelity as follows: ； in, Indicates the absolute value, represents the network super-resolution reconstructed image, represents the true label image; Structural similarity loss as follows: ； ； in, and Represents images respectively and images The average value of and They represent the variance of the two, represents the covariance between the two, and is a constant less than 0.003; Representing an image and images The structural similarity of Representing an image and images Structural similarity loss; Spatial gradient loss as follows: ； ； ； ； ； in, , Represents the Sobel operator in the horizontal and vertical directions respectively, Represents a splicing operation, To predict the image Sobey operator and The concatenated result of sliding multiplication represents the predicted image Spatial gradient changes, For real images Sobey operator and The result of sliding multiplication represents the real image Spatial gradient changes, The total losses are as follows: ； in, represents the total loss, 、 、 represents a hyperparameter.

7. A DW image super-resolution reconstruction system, characterized in that: include: The multi-directional projection fusion module is used to calculate the similarity between the target direction and other directions through cosine similarity in Q space, select the DW image slices in the N directions with the largest cosine similarity to form a neighboring image, and use the RDN network architecture to extract the features of the target direction and the N neighboring direction images respectively; the target direction features and the features of the N neighboring direction images are orthogonally projected and fused to generate fused features; An implicit neural network module is used to combine the fusion features with the high-resolution image coordinates, map them into a continuous function, and use an image-to-image learning method to add data fidelity loss, spatial gradient change loss, and structural similarity loss to reconstruct super-resolution DW images of any scale; The loss calculation module is used to calculate data fidelity loss, spatial gradient change loss, and structural similarity loss, and optimize the model based on these losses.

8. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory is used to store instructions, and the processor is used to execute the instructions to implement the DW image super-resolution reconstruction method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the DW image super-resolution reconstruction method according to any one of claims 1 to 6 is implemented.

Images