Sparse Laplacian field image reconstruction method based on wavelet guided CNN (Convolutional Neural Network)

Abstract

The invention belongs to the technical field of image processing, and discloses a sparse Laplacian field image reconstruction method based on wavelet guided CNN, comprising: acquiring an RGB image, and performing spatial convolution on the RGB image and a discrete Laplacian operator kernel to obtain a Laplacian field image; based on the Laplacian field image, reconstructing an image from the sparse Laplacian field by using a wavelet-guided convolutional neural network to obtain a reconstructed RGB image; and outputting the reconstructed RGB image. According to the method, the sparse Laplacian field is provided for representing the image, the Poisson equation is solved in cooperation with the wavelet-guided convolutional neural network, and the precision and efficiency of image reconstruction are improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a sparse Laplacian field image reconstruction method based on wavelet-guided CNN. Background Technology

[0002] Image processing techniques mainly include gradient domain image processing methods, traditional Poisson solution methods, and deep learning methods. However, these image processing methods face the following drawbacks: 1) Gradient domain image processing: It relies on the gradient field to solve the Poisson equation, which requires storing and processing two-dimensional vector fields (gradient x / y components), resulting in high computational complexity and low storage efficiency.

[0003] 2) Traditional Poisson solver: Direct methods (e.g., Fast Fourier Transform / Discrete Cosine Transform): have a computational complexity of O(N log N), making them difficult to process high-resolution images in real time.

[0004] Iterative methods (e.g., Jacobi iteration, multigrid method): require multiple iterations to converge, resulting in poor real-time performance.

[0005] Sensitive to noise, traditional methods are easily affected by noise interference, leading to reconstruction failure.

[0006] 3) Deep learning methods: It has a large number of parameters (e.g., it requires 2 hours to preprocess a 256×256 image) and poor generalization ability (it requires a fixed input resolution), making it unable to flexibly adapt to different scenarios.

[0007] Therefore, existing image reconstruction techniques still suffer from low storage efficiency, high computational complexity, and poor generalization ability, and the techniques need further improvement. Summary of the Invention

[0008] The technical problem to be solved by this invention is to provide a sparse Laplacian field image reconstruction method based on wavelet-guided CNN to address the shortcomings of existing image reconstruction techniques, such as low storage efficiency, high computational complexity, and poor generalization ability.

[0009] The technical solution adopted by this invention to solve the technical problem is as follows: In a first aspect, the present invention provides a sparse Laplacian field image reconstruction method based on wavelet-guided CNN, comprising: The RGB image is obtained by spatially convolving the RGB image with a discrete Laplacian operator kernel to obtain the Laplacian field image. Based on the Laplacian field image, a wavelet-guided convolutional neural network is used to reconstruct the image from the sparse Laplacian field to obtain the reconstructed RGB image. Output the reconstructed RGB image.

[0010] In one implementation, obtaining the Laplacian field image by spatially convolving the RGB image with a discrete Laplacian operator kernel includes: The discrete Laplacian operator kernel is slid across the RGB image, a weighted sum is calculated at each position, and a ring of zeros is added outside the boundaries of the RGB image using a zero-padding method to keep the output size the same as the input.

[0011] In one implementation, the discrete Laplace operator kernel is: .

[0012] In one implementation, the formula for calculating the sparse Laplace field is: ; Where i and j represent the row and column indices of the output position, respectively; c represents the channel index; Indicates the input image In position The value of channel c; if it is out of bounds, it is filled with 0; Nuclear Index This indicates the row and column corresponding to the kernel.

[0013] In one implementation, the wavelet-guided convolutional neural network includes: a convolution module, a downsampling module, an upsampling module, and a summing module; wherein the convolution kernels of the convolution module, the downsampling module, and the upsampling module are globally shared convolution kernels.

[0014] In one implementation, the step of reconstructing the image from the sparse Laplacian field using a wavelet-guided convolutional neural network based on the Laplacian field image to obtain the reconstructed RGB image includes: The Laplacian field image is input into the wavelet-guided convolutional neural network; For the input Laplacian field image, fill the image with a boundary of width 1 and all pixel values ​​of 0; The Laplacian field image with filled boundaries is downsampled and upsampled with multiple convolution kernels, and a 1-pixel ring around the upsampled image is cropped to obtain the reconstructed RGB image.

[0015] In one implementation, downsampling and upsampling the boundary-filled Laplacian field image with multiple convolutional kernels includes: The Laplacian field image L1 after filling the boundary is convolved with the convolution kernel H and then downsampled to obtain the Laplacian field image L2; Repeat the convolution and downsampling operations until the Laplacian field image L is obtained. n ; The Laplace field image L n After convolution with kernel G, image U is obtained. n ; The image U n First, upsample the image. Then, convolve the upsampled points with kernel K to obtain the first convolution result. Next, convolve the Laplacian field image with kernel G to obtain the second convolution result. Finally, add the first convolution result to the second convolution result to obtain the image U. n-1 ; Repeat the convolution and upsampling operations to obtain the upsampled image.

[0016] Secondly, the present invention provides a sparse Laplacian field image reconstruction system based on wavelet-guided CNN, comprising: The sparse Laplacian field generation module is used to acquire RGB images and obtain Laplacian field images by spatially convolving the RGB images with discrete Laplacian operator kernels. The image reconstruction module is used to reconstruct the image from the sparse Laplacian field based on the Laplacian field image using a wavelet-guided convolutional neural network, and obtain the reconstructed RGB image. The output module is used to output the reconstructed RGB image.

[0017] Thirdly, the present invention provides a terminal comprising: a processor and a memory, the memory storing a sparse Laplacian field image reconstruction program based on wavelet-guided CNN, wherein when the processor executes the sparse Laplacian field image reconstruction program based on wavelet-guided CNN, it is used to implement the operation of the sparse Laplacian field image reconstruction method based on wavelet-guided CNN as described in the first aspect.

[0018] Fourthly, the present invention also provides a computer-readable storage medium storing a sparse Laplacian field image reconstruction program based on wavelet-guided CNN, wherein the sparse Laplacian field image reconstruction program based on wavelet-guided CNN, when executed by a processor, is used to implement the operation of the sparse Laplacian field image reconstruction method based on wavelet-guided CNN as described in the first aspect.

[0019] The present invention, by employing the above technical solution, has the following effects: This invention acquires an RGB image, generates a sparse Laplacian field using a discrete Laplacian operator kernel and the RGB image, and obtains a Laplacian field image. Based on the Laplacian field image, a wavelet-guided convolutional neural network is used to reconstruct the image from the sparse Laplacian field, obtaining the reconstructed RGB image. The reconstructed RGB image is then output. This invention proposes a sparse Laplacian field representation of the image, combined with a wavelet-guided convolutional neural network to solve the Poisson equation, improving the accuracy and efficiency of image reconstruction. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the sparse Laplacian field image reconstruction method based on wavelet-guided CNN in this invention.

[0022] Figure 2 This is a schematic diagram of the pixel value distribution of each image in the BSDS dataset of this invention.

[0023] Figure 3 This is a schematic diagram of the pixel value distribution of each image in the DIV2K dataset of this invention.

[0024] Figure 4 This is a flowchart of the reconstruction of the Laplace field using a wavelet-guided convolutional neural network in this invention.

[0025] Figure 5 This is a numerical visualization diagram of each convolution kernel in this invention.

[0026] Figure 6 This is a schematic diagram comparing the recovery effects of the method provided in this invention and the wavelet method.

[0027] Figure 7 This is a functional schematic diagram of the terminal in one implementation of the present invention.

[0028] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0030] Exemplary methods Image processing techniques mainly include gradient domain image processing methods, traditional Poisson solution methods, and deep learning methods. However, these image processing methods face the following drawbacks: 1) Gradient domain image processing: It relies on the gradient field to solve the Poisson equation, which requires storing and processing two-dimensional vector fields (gradient x / y components), resulting in high computational complexity and low storage efficiency.

[0031] 2) Traditional Poisson solver: Direct methods (e.g., Fast Fourier Transform / Discrete Cosine Transform): have a computational complexity of O(N log N), making them difficult to process high-resolution images in real time.

[0032] Iterative methods (e.g., Jacobi iteration, multigrid method): require multiple iterations to converge, resulting in poor real-time performance.

[0033] Sensitive to noise, traditional methods are easily affected by noise interference, leading to reconstruction failure.

[0034] 3) Deep learning methods: It has a large number of parameters (e.g., it requires 2 hours to preprocess a 256×256 image) and poor generalization ability (it requires a fixed input resolution), making it unable to flexibly adapt to different scenarios.

[0035] Therefore, existing image reconstruction techniques still suffer from low storage efficiency, high computational complexity, and poor generalization ability, and the techniques need further improvement.

[0036] To address the above technical problems, this invention provides a sparse Laplacian field image reconstruction method based on wavelet-guided CNN, comprising: acquiring an RGB image; spatially convolving the RGB image with a discrete Laplacian operator kernel to obtain a Laplacian field image; reconstructing the image from the sparse Laplacian field using a wavelet-guided convolutional neural network based on the Laplacian field image to obtain a reconstructed RGB image; and outputting the reconstructed RGB image. This invention proposes a sparse Laplacian field representation of the image, combined with a wavelet-guided convolutional neural network to solve the Poisson equation, thereby improving the accuracy and efficiency of image reconstruction.

[0037] like Figure 1 As shown, this embodiment of the invention provides a sparse Laplacian field image reconstruction method based on wavelet-guided CNN, including the following steps: Step S100: Obtain the RGB image and obtain the Laplacian field image by spatially convolving the RGB image with the discrete Laplacian operator kernel.

[0038] In this embodiment, a sparse Laplacian field image reconstruction method based on wavelet-guided CNN (convolutional neural network) is proposed. This method is an image reconstruction method based on sparse Laplacian field and wavelet-guided convolutional neural network. The basic scheme of this image reconstruction method is as follows: We propose a sparse Laplacian field representation of the image, and use wavelet-guided convolutional neural networks to solve the Poisson equation: 1. Extract the sparse scalar field of an image using the Laplacian operator (only the original data is needed). Figure 1 / 2 storage); 2. Design a lightweight CNN network (wavelet-guided convolutional neural network with a multi-scale structure using shared convolutional kernels) to reconstruct images, supporting input images of arbitrary resolution.

[0039] Specifically, in one implementation of this embodiment, step S100 includes the following steps: Step S101: Slide the discrete Laplacian operator kernel across the RGB image, calculate the weighted sum at each position, and use the zero-padding method to add a ring of zero values ​​outside the boundary of the RGB image to keep the output size the same as the input.

[0040] In this embodiment, an RGB three-channel color image u is first acquired and used as the image to be reconstructed. Then, the discrete Laplacian operator kernel is used. Compared with the original image The Laplace field is obtained by convolution calculation. : ; Specifically, using the discrete Laplace operator kernel Compared with the original image The Laplace field is obtained by convolution calculation. The process is as follows: Mathematically, an image u can be represented by a (w*h*3 matrix, where w is the width of the image and h is the height of the image).

[0041] Example of Laplace field calculation: In image processing, the Laplacian field is calculated by applying the discrete Laplacian operator (a second-order derivative approximation) to an image and is commonly used for edge detection. Image u is an RGB three-channel color image, represented as a matrix with width w (horizontal) and height h (vertical), with dimensions w×h×3. Each channel (red, green, blue) is processed independently. k is a 3×3 convolution kernel, as shown above. The convolution operation involves sliding the kernel across the image and calculating a weighted sum (dot product) at each location. Since the kernel size is 3×3, boundary processing is necessary. Zero-padding is used here: a ring of zeros is added around the image boundaries to keep the output size the same as the input (i.e., w×h).

[0042] The formula for calculating the above sparse Laplace field is as follows: ; Where i and j represent the row and column indices of the output position (starting from 0), respectively. c represents the channel index (0 for red, 1 for green, 2 for blue). Indicates the position of the input image. The value of channel c; if it is out of bounds, it is filled with 0; Nuclear Index Indicates the row and column of the corresponding kernel (e.g., (Corresponding to the upper left corner of the nucleus).

[0043] In this embodiment, a Laplacian field image is obtained by spatially convolving an RGB image with a discrete Laplacian operator kernel. The Laplacian field image is then input into a wavelet-guided convolutional neural network, which is used to reconstruct the image from the sparse Laplacian field, thus obtaining the reconstructed RGB image.

[0044] To verify the sparsity of the Laplace field described above, the sparsity is verified in the following way: First, select the BSDS dataset (Berkeley Segmentation Data Set) and the DIV2K dataset (DIVerse 2K resolution high quality images, a high-resolution image dataset containing high-quality image data at 2K resolution, often used in deep learning and computer vision research). In two different datasets, the grayscale histograms of each image are calculated separately and then overlaid for display, as shown below. Figure 2 and Figure 3 As shown, Figure 2 This is a schematic diagram illustrating the pixel value distribution of each image in the BSDS dataset. Figure 3 This is a schematic diagram illustrating the pixel value distribution of each image in the DIV2K dataset. Figure 2 and Figure 3 In the graph, the horizontal axis represents the grayscale value of a pixel, and the vertical axis represents the logarithmic value of the frequency. The red curve represents the mean. By overlaying the graph, it can be seen that the vast majority of pixels in the Laplacian field have values ​​of zero or close to zero, thus conforming to the sparsity property.

[0045] Specifically, the selected BSDS dataset contains 500 images with a resolution of 481*321 or 321*481. In this embodiment, they are divided into two groups of 400 and 100 images, then cropped to 256*256 and rotated, resulting in a training set of 6400 images and a test set of 1600 images.

[0046] The selected DIV2K dataset contains 1000 high-resolution images. The original resolution is not fixed; in this embodiment, they are all cropped to 1024*1024, resulting in a training set of 1600 images and a test set of 202 images.

[0047] For the BSDS and DIV2K datasets mentioned above, the gray-level histogram for each image is calculated. Specifically, the gray-level histogram calculation method is as follows: For each image u, the Laplace transform is performed on the image. For each pixel in the image, there are three brightness channels. Here, the average of the three channels is taken as its brightness value: a) Calculate each image The grayscale histogram is obtained and normalized to a probability density (the sum of the areas is 1). b) Calculate the coordinates of each bar in the histogram and represent them as a scatter plot; c) Repeat the above process for each image, and overlay the scatter plots of all the images to form a scatter plot of pixel brightness distribution of all images in the dataset. d) Take the logarithm of the vertical axis scale of the image for easier viewing. The final result is as follows: Figure 2 and Figure 3 As shown.

[0048] The Laplace field generated in this embodiment is sparse, satisfies a stable distribution, and has a unique solution by solving the Poisson equation. This uniqueness is mathematically guaranteed, providing a theoretical basis for the method proposed in this embodiment. Moreover, the Poisson equation for a given Laplace field has a unique solution with appropriate boundary conditions, ensuring the uniqueness of the Poisson equation solution.

[0049] like Figure 1 As shown, this embodiment of the invention provides a sparse Laplacian field image reconstruction method based on wavelet-guided CNN, including the following steps: Step S200: Based on the Laplacian field image, a wavelet-guided convolutional neural network is used to reconstruct the image from the sparse Laplacian field to obtain the reconstructed RGB image; Step S300: Output the reconstructed RGB image.

[0050] In this embodiment, for the Laplacian field image obtained above, the Laplacian field image is input into a wavelet-guided convolutional neural network, and the image is reconstructed from the sparse Laplacian field using the wavelet-guided convolutional neural network to obtain the reconstructed RGB image.

[0051] As an example, the wavelet-guided convolutional neural network used in this embodiment includes: a convolution module, a downsampling module, an upsampling module, and an summing module; wherein the convolution kernels of the convolution module, the downsampling module, and the upsampling module are globally shared convolution kernels.

[0052] Specifically, in one implementation of this embodiment, step S200 includes the following steps: Step S201: Input the Laplacian field image into the wavelet-guided convolutional neural network; Step S202: For the input Laplacian field image, fill the area around the image with a boundary of width 1 and all pixel values ​​of 0. Step S203: The Laplacian field image after filling the boundary is downsampled and upsampled with multiple convolution kernels, and a ring of pixels with a width of 1 is cropped around the upsampled image to obtain the reconstructed RGB image.

[0053] like Figure 4 As shown, in this embodiment, Figure 4 This describes the process of using a wavelet-guided convolutional neural network to reconstruct an image from a sparse Laplacian field. The red shaded area represents the input Laplacian field. The decomposition process (analysis process in wavelet transform), the green shaded area is the output. The reconstruction process (synthesis process in wavelet transform).

[0054] During the decomposition process, input 1 and convolution kernel (Corresponding to a low-pass filter in wavelets) Convolve, then downsample to the next scale level. Repeat this process until the final scale is reached. Then, With convolution kernel Perform convolution to obtain a reconstructed image. .

[0055] During the reconstruction process, the image Upsampled and convolved with kernel Perform convolution. Meanwhile, n -1 and a convolutional kernel (Corresponding to a high-pass filter in wavelets) Convolution is performed. The convolution kernels (H, G, and K) are learned and shared at each stage, thus yielding fewer than 200 parameters regardless of the input resolution. Furthermore, its computational complexity is linearly related to the input pixels, resulting in high computational performance.

[0056] In one implementation of this embodiment, step S203 includes the following steps: Step S203a: Convolve the Laplacian field image L1 after filling the boundary with the convolution kernel H and then downsample to obtain the Laplacian field image L2; Step S203b: Repeat the convolution and downsampling operations until the Laplacian field image L is obtained. n ; Step S203c, the Laplacian field image L n After convolution with kernel G, image U is obtained. n ; Step S203d, the image U n First, upsample the image. Then, convolve the upsampled points with kernel K to obtain the first convolution result. Next, convolve the Laplacian field image with kernel G to obtain the second convolution result. Finally, add the first convolution result to the second convolution result to obtain the image U. n-1 ; Step S203e: Repeat the convolution and upsampling operations to obtain the upsampled image.

[0057] In this embodiment, the specific process of reconstructing the image from the sparse Laplacian field using a wavelet-guided convolutional neural network is as follows: S1. For the input Laplacian field image, first fill the image with a boundary of width 1 and all pixel values ​​of 0 to obtain image L1, which facilitates subsequent convolution calculations. For example, for The image, after being filled, becomes... .

[0058] S2. L1 and kernel H are convolved and then downsampled to obtain L2. Downsampling takes a value every other pixel in both the height and width dimensions. This is equivalent to reducing the image resolution to half its original value in both the height and width directions. After the operation, the image size becomes: height = h / 2, width = w / 2 (ignoring the padding).

[0059] S3. Repeat the above steps to obtain L3, L4, and so on until L... n The downsampling is performed the maximum of the height and width of the input image, i.e., max(h, w). The base-2 logarithm of this maximum value is calculated, rounded up (ceil), and then subtracted by one. Finally, the result is converted to an integer.

[0060] .

[0061] S4. L n After convolution with kernel G, image U is obtained. n .

[0062] S5. Next, upsample from the lowest layer: Image U n First, upsample the data. Then, convolve the upsampled data with kernel K. Finally, combine this result (the first convolution result) with kernel L. n-1 Adding the result of the convolution with kernel G (the result of the second convolution) yields U. n-1 Upsampling is an interpolation process of the image matrix, placing pixel values ​​at even-numbered index positions and keeping odd-numbered index positions at zero, thus achieving a 2x upsampling (doubling the resolution).

[0063] S6. Repeat this process continuously and upsample to obtain the result U1.

[0064] S7. Finally, crop a 1-pixel ring around U1 to obtain the input image, that is, restore the original image and obtain the reconstructed RGB image.

[0065] In this embodiment, the convolution kernel can adaptively change its size. When set to its minimum, the corresponding number of parameters is minimized. Both convolution kernels H and G are 5x5x3 matrices, and convolution kernel K is a 3x3x3 matrix. The numerical values ​​can be visualized as follows: Figure 5 As shown.

[0066] like Figure 6 As shown in the figure, multiple experimental results indicate that the method provided in this embodiment has a faster computation speed. Table 1 below shows the time required for several different methods to recover the Laplacian domain on images with different resolutions. It can be seen that the time used by this method is similar to that of the wavelet method, and much less than that of the traditional DCT method.

[0067] Table 1. Time required to recover the Laplacian domain on images at different resolutions using multiple different methods:

[0068] Furthermore, the recovery accuracy of this method is much higher than that of the wavelet method, as shown in Table 2 below.

[0069] Table 2. Recovery accuracy of multiple different methods on images at different resolutions:

[0070] In this embodiment, a wavelet-guided convolutional neural network is used to reconstruct an image from a sparse Laplacian field, resulting in a reconstructed RGB image. A lightweight network is then used to solve the Poisson equation, overcoming the limitations of traditional iterative methods (such as FFT) and deep learning solvers. Furthermore, by designing convolutional kernels H, G, and K to be reused across all layers, the number of parameters is reduced to as low as 177, achieving cross-scale kernel sharing. This embodiment does not require a predefined wavelet basis; instead, the optimal filter is learned through training, implementing a data-driven wavelet basis scheme.

[0071] Based on the above technical solution, this embodiment allows for adjustable convolution kernel channels (8 / 16 / 32 channels) to improve scene adaptability. In the case of the 32-channel model, the PSNR reaches 61.58dB, and the error is reduced by 100 times compared to the baseline, which can further improve the reconstruction accuracy.

[0072] As an alternative to the technical solution of this embodiment, the Laplacian operator can be replaced with a biharmonic operator or a Hessian matrix for convolution calculation; the shared convolution kernel can be removed and the wavelet-guided convolutional neural network can be modified; or a hybrid solution scheme (CNN + traditional iteration) can be used to solve the Poisson equation.

[0073] This embodiment achieves the following technical effects through the above technical solution: This embodiment acquires an RGB image, generates a sparse Laplacian field using a discrete Laplacian operator kernel and the RGB image, thus obtaining a Laplacian field image. Based on the Laplacian field image, a wavelet-guided convolutional neural network is used to reconstruct the image from the sparse Laplacian field, yielding the reconstructed RGB image, which is then output. This embodiment proposes a sparse Laplacian field representation of the image, combined with a wavelet-guided convolutional neural network to solve the Poisson equation, improving the accuracy and efficiency of image reconstruction.

[0074] Exemplary device Based on the above embodiments, the present invention also provides a sparse Laplacian field image reconstruction system based on wavelet-guided CNN, comprising: The sparse Laplacian field generation module is used to acquire RGB images and obtain Laplacian field images by spatially convolving the RGB images with discrete Laplacian operator kernels. The image reconstruction module is used to reconstruct the image from the sparse Laplacian field based on the Laplacian field image using a wavelet-guided convolutional neural network, and obtain the reconstructed RGB image. The output module is used to output the reconstructed RGB image.

[0075] This embodiment achieves the following technical effects through the above technical solution: This embodiment acquires an RGB image, generates a sparse Laplacian field using a discrete Laplacian operator kernel and the RGB image, thus obtaining a Laplacian field image. Based on the Laplacian field image, a wavelet-guided convolutional neural network is used to reconstruct the image from the sparse Laplacian field, yielding the reconstructed RGB image, which is then output. This embodiment proposes a sparse Laplacian field representation of the image, combined with a wavelet-guided convolutional neural network to solve the Poisson equation, improving the accuracy and efficiency of image reconstruction.

[0076] Based on the above embodiments, the present invention also provides a terminal, the principle block diagram of which can be as follows: Figure 7 As shown.

[0077] The terminal includes: a processor, a memory, an interface, a display screen, and a communication module connected via a system bus; wherein, the processor of the terminal provides computing and control capabilities; the memory of the terminal includes a computer-readable storage medium and internal memory; the computer-readable storage medium stores an operating system and computer programs; the internal memory provides an environment for the operation of the operating system and computer programs in the computer-readable storage medium; the interface is used to connect to external devices; the display screen is used to display relevant information; and the communication module is used to communicate with a cloud server or other devices.

[0078] When executed by the processor, this computer program is used to implement the operation of a sparse Laplacian field image reconstruction method based on wavelet-guided CNN.

[0079] It will be understood by those skilled in the art that Figure 7 The schematic diagram shown is merely a partial structural diagram related to the present invention and does not constitute a limitation on the terminal to which the present invention is applied. A specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0080] In one embodiment, a terminal is provided, comprising: a processor and a memory, the memory storing a sparse Laplacian field image reconstruction program based on wavelet-guided CNN, the sparse Laplacian field image reconstruction program based on wavelet-guided CNN being executed by the processor to implement the operation of the sparse Laplacian field image reconstruction method based on wavelet-guided CNN as described above.

[0081] In one embodiment, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a sparse Laplacian field image reconstruction program based on wavelet-guided CNN, which, when executed by a processor, is used to implement the operation of the sparse Laplacian field image reconstruction method based on wavelet-guided CNN as described above.

[0082] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, database, or other media used in the embodiments provided by this invention can include both non-volatile and volatile memory.

[0083] In summary, this invention provides a sparse Laplacian field image reconstruction method based on wavelet-guided CNN, comprising: acquiring an RGB image; obtaining a Laplacian field image by spatially convolving the RGB image with a discrete Laplacian operator kernel; reconstructing the image from the sparse Laplacian field using a wavelet-guided convolutional neural network based on the Laplacian field image to obtain the reconstructed RGB image; and outputting the reconstructed RGB image. This invention proposes a sparse Laplacian field representation of the image, combined with a wavelet-guided convolutional neural network to solve the Poisson equation, thereby improving the accuracy and efficiency of image reconstruction.

[0084] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A sparse Laplacian field image reconstruction method based on wavelet-guided CNN, characterized in that, include: The RGB image is obtained by spatially convolving the RGB image with a discrete Laplacian operator kernel to obtain the Laplacian field image. Based on the Laplacian field image, a wavelet-guided convolutional neural network is used to reconstruct the image from the sparse Laplacian field to obtain the reconstructed RGB image. Output the reconstructed RGB image.

2. The sparse Laplacian field image reconstruction method based on wavelet-guided CNN according to claim 1, characterized in that, The step of obtaining a Laplacian field image by spatially convolving an RGB image with a discrete Laplacian operator kernel includes: The discrete Laplacian operator kernel is slid across the RGB image, a weighted sum is calculated at each position, and a ring of zeros is added outside the boundaries of the RGB image using a zero-padding method to keep the output size the same as the input.

3. The sparse Laplacian field image reconstruction method based on wavelet-guided CNN according to claim 2, characterized in that, The discrete Laplace operator kernel is: 。 4. The sparse Laplacian field image reconstruction method based on wavelet-guided CNN according to claim 2, characterized in that, The formula for calculating the sparse Laplace field is as follows: ； Where i and j represent the row and column indices of the output position, respectively; c represents the channel index; Indicates the input image In position The value of channel c; if it is out of bounds, it is filled with 0; Nuclear Index This indicates the row and column corresponding to the kernel.

5. The sparse Laplacian field image reconstruction method based on wavelet-guided CNN according to claim 1, characterized in that, The wavelet-guided convolutional neural network includes a convolution module, a downsampling module, an upsampling module, and an summing module; wherein the convolution kernels of the convolution module, the downsampling module, and the upsampling module are globally shared convolution kernels.

6. The sparse Laplacian field image reconstruction method based on wavelet-guided CNN according to claim 1, characterized in that, The process of reconstructing an image from the sparse Laplacian field image using a wavelet-guided convolutional neural network to obtain the reconstructed RGB image includes: The Laplacian field image is input into the wavelet-guided convolutional neural network; For the input Laplacian field image, fill the image with a boundary of width 1 and all pixel values ​​of 0; The Laplacian field image with filled boundaries is downsampled and upsampled with multiple convolution kernels, and a 1-pixel ring around the upsampled image is cropped to obtain the reconstructed RGB image.

7. The sparse Laplacian field image reconstruction method based on wavelet-guided CNN according to claim 6, characterized in that, The step of downsampling and upsampling the Laplacian field image after filling the boundaries with multiple convolutional kernels includes: The Laplacian field image L1 after filling the boundary is convolved with the convolution kernel H and then downsampled to obtain the Laplacian field image L2; Repeat the convolution and downsampling operations until the Laplacian field image L is obtained. n ; The Laplace field image L n After convolution with kernel G, image U is obtained. n ; The image U n First, upsample the image. Then, convolve the upsampled points with kernel K to obtain the first convolution result. Next, convolve the Laplacian field image with kernel G to obtain the second convolution result. Finally, add the first convolution result to the second convolution result to obtain the image U. n-1 ; Repeat the convolution and upsampling operations to obtain the upsampled image.

8. A sparse Laplacian field image reconstruction system based on wavelet-guided CNN, characterized in that, include: The sparse Laplacian field generation module is used to acquire RGB images and obtain Laplacian field images by spatially convolving the RGB images with discrete Laplacian operator kernels. The image reconstruction module is used to reconstruct the image from the sparse Laplacian field based on the Laplacian field image using a wavelet-guided convolutional neural network, and obtain the reconstructed RGB image. The output module is used to output the reconstructed RGB image.

9. A terminal, characterized in that, include: The processor and memory, wherein the memory stores a sparse Laplacian field image reconstruction program based on wavelet-guided CNN, the sparse Laplacian field image reconstruction program based on wavelet-guided CNN being executed by the processor to implement the operation of the sparse Laplacian field image reconstruction method based on wavelet-guided CNN as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a sparse Laplacian field image reconstruction program based on wavelet-guided CNN, which, when executed by a processor, is used to implement the operation of the sparse Laplacian field image reconstruction method based on wavelet-guided CNN as described in any one of claims 1-7.

Images