Image compressed sensing reconstruction method and system based on recursive diffusion model

Abstract

The invention discloses an image compressed sensing reconstruction method and system based on a recursive diffusion model, and belongs to the technical field of image processing and compressed sensing, and the method comprises the steps: obtaining a to-be-reconstructed original image; performing block compressed sensing sampling on an original image to be reconstructed to obtain a complete observation value; initializing the complete observation value by adopting a pseudo-inverse reprojection operation to obtain an initial reconstructed image; performing complete reconstruction on the initial reconstruction image by using a recursive diffusion model; wherein in the whole image domain, the initial reconstruction image is used as the image estimation of the current iteration, through a lightweight recursion UNet submodule and an operator condition circulation prior submodule in the recursion diffusion model, multi-step iteration reconstruction is carried out, and a final reconstruction image is output. According to the method, a recursive refinement mechanism and stride memory prior are introduced, high-quality image reconstruction is realized under a small number of iteration steps, and the calculation and storage overhead is remarkably reduced.

Description

Technical Field

[0001] This invention belongs to the field of image processing and compressed sensing technology, and particularly relates to an image compressed sensing reconstruction method and system based on a recursive diffusion model. Background Technology

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

[0003] Since its inception, compressed sensing (CS) has attracted widespread attention in the field of signal acquisition and reconstruction. Its core advantage lies in the fact that if a signal exhibits sparsity in a certain transform domain, it can achieve accurate reconstruction of the original signal at a sampling rate far lower than the Nyquist rate. This characteristic makes compressed sensing irreplaceable in resource-constrained scenarios such as medical imaging, remote sensing monitoring, magnetic resonance imaging, and single-pixel imaging.

[0004] However, despite the increasing sophistication of compressed sensing theory, reconstructing the original image with the highest possible quality from a limited set of observations remains a core research hotspot and challenge in the field. Traditional compressed sensing reconstruction methods, such as the Iterative Soft Thresholding Algorithm (ISTA), Basis Pursuit, and their variants, while theoretically guaranteeing reconstruction, generally suffer from high computational complexity, poor robustness to noise, and slow reconstruction speed in practical applications, making it difficult to meet the demands for real-time and high-quality reconstruction.

[0005] In recent years, deep learning methods, represented by Convolutional Neural Networks (CNNs), have demonstrated significant advantages in the field of compressed sensing of images, greatly improving the quality and efficiency of image reconstruction. Currently, deep learning-based compressed sensing reconstruction methods are mainly divided into two categories: First, end-to-end compressed sensing reconstruction networks, such as CSNet (Compressive Sensing Network) and SCSNet (Structured Compressive Sensing Network). These networks can significantly reduce computational complexity and achieve reconstruction accuracy superior to traditional methods during the reconstruction process. However, these models rely excessively on fixed network architectures, have limited generalization ability, and are difficult to adapt to the reconstruction needs of various image types or complex scenes. They also suffer from the problem of uninterpretable model decision-making processes. Second, deeply unfolded compressed sensing reconstruction networks, such as LISTA (Learned Iterative Shrinkage Thresholding Algorithm) and ISTA-Net+ (Iterative Shrinkage Thresholding Algorithm Networkplus). Networks such as the Iterative Shrink Thresholding Algorithm Network Enhancement Edition (ITTA, Alternating Direction Multiplier Method ADMM) expand the iterative process of iterative algorithms (such as the Iterative Soft Thresholding Algorithm ISTA, Alternating Direction Multiplier Method ADMM, etc.) into learnable neural network layers. While inheriting the interpretability and stability of traditional algorithms, they accelerate the reconstruction process. However, these networks are still limited by a fixed iterative framework, have high computational redundancy, and their reconstruction effect drops significantly in low sampling rate or high noise scenarios, making it difficult to meet the needs of practical applications.

[0006] Diffusion models, as emerging generative models, have performed exceptionally well in image processing tasks such as image generation, denoising, and super-resolution. They achieve high-quality image reconstruction through an iterative process of forward denoising and reverse denoising. However, existing compressed sensing methods based on diffusion models, such as DDNM (Denoising Diffusion Null-Space Model) and IDM (Invertible Diffusion Models), typically unfold the T-step diffusion process into T independent U-Net structures. This results in a huge number of model parameters and extremely high computational costs, severely limiting their practical application in resource-constrained scenarios. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, this invention proposes an image compressed sensing reconstruction method and system based on a recursive diffusion model. By introducing a recursive refinement mechanism and a step-by-step memory prior, high-quality image reconstruction is achieved with a small number of iterations, significantly reducing computational and storage overhead. This reduces the computational complexity and number of parameters while maintaining the high reconstruction quality of the diffusion model, effectively improving the model's practicality and generalization ability.

[0008] In the first aspect, this invention proposes an image compressed sensing reconstruction method based on a recursive diffusion model.

[0009] A compressed sensing reconstruction method for images based on a recursive diffusion model includes: Obtain the original image to be reconstructed; The original image to be reconstructed is subjected to block-based compressed sensing sampling to obtain complete observations; A pseudo-reverse projection operation is used to initialize the complete observations to obtain the initial reconstructed image; The recursive diffusion model is used to fully reconstruct the initial image. In the full image domain, the initial reconstructed image is used as the image estimate for the current iteration. Multi-step iterative reconstruction is performed through the lightweight recursive UNet submodule and the operator conditional loop prior submodule in the recursive diffusion model to output the final reconstructed image.

[0010] A further technical solution involves performing block-based compressed sensing sampling on the original image to be reconstructed to obtain complete observations, including: Press the original image The blocks are divided in a non-overlapping manner. After each block is expanded, it is sampled using the learned measurement matrix to obtain the observation values ​​for each block, as follows: ; In the above formula, This represents the first segment after the original image is divided into blocks. piece, Indicates to The sampling matrix used during sampling Yes The corresponding observations obtained after sampling; By combining the observations from all blocks, a complete set of observations can be obtained. .

[0011] A further technical solution involves using a pseudo-reverse projection operation to initialize the complete observations, resulting in an initial reconstructed image, including: During the initialization phase, a pseudo-reverse projection is performed on each block and then collapsed back to the entire image to obtain an initial image estimate for the entire map domain. ,for: ; In the above formula, Indicates will The complete observations after being combined The pseudo-inverse of the sampling matrix. It is the initial image estimate for the entire map domain, i.e., the initial reconstructed image.

[0012] A further technical solution involves a multi-step iterative image reconstruction process: Based on the image estimation of the current iteration, combined with complete observations, physical guidance features are constructed. The image estimation and physical guidance features of the current iteration are input into the operator conditional loop prior submodule to generate the prior image and adaptive fusion gate for the current iteration; The image estimate of the current iteration is input into the lightweight recursive UNet submodule, which extracts the intermediate features of the current image estimate. Then, it is refined through multiple rounds of residual refinement by a recursive unit with shared weights to obtain refined features. The refined features are restored to obtain noise prediction. The predicted noise is then denoised analytically to obtain the current image estimate after noise removal. Based on the adaptive fusion gate, the prior image and the current image estimate after denoising are adaptively fused, and the data consistency projection alignment measurement constraint is executed. Then, the image estimate for the next iteration is obtained by updating through DDIM. The process is iterated continuously until the set number of iterations is reached, and the final image estimate is used as the final reconstructed image and then output.

[0013] A further technical solution utilizes a recursive diffusion model for multi-step iterative reconstruction, with the loss function being a weighted sum of pixel-level mean square error loss, multi-scale mean square error loss, measurement consistency loss, and projection consistency loss.

[0014] A further technical solution involves inputting the image estimate from the current iteration into a lightweight recursive UNet submodule to obtain the noise-removed current image estimate, including: The current image is estimated in the pixel downsampling space, and then the encoder-decoder and the cross-layer skip connection UNet backbone network are used in conjunction with residual blocks to perform basic reconstruction, outputting the intermediate features of the pixel downsampling space. A recursive unit with shared weights is introduced to perform multi-round residual refinement of intermediate features, as follows: each round recursively applies the features from the previous round. As input, the residual enhancement amount is generated. The current wheel feature is formed through the residual loop. The intermediate features are used as the features of the initial round; after... After recursively applying the wheel residuals, the final refined features are obtained; The refined features are restored to the original pixel space using PixelShuffle×2 to obtain the noise prediction. ; Based on the predicted noise Perform analytical denoising as follows: ; In the above formula, It is a noisy image at step t. It is noise predicted by Lightweight Recursive UNet (i.e., RNR-UNet). It is the cumulative attenuation factor in noise scheduling. It is the current image estimate after noise removal in the current round.

[0015] A further technical solution, the formula for DDIM updates, is: ; In the above formula, It is noise predicted by RNR-UNet. It is the current image estimate after noise removal in the current round. This is the sampling result at the next time step. It is the cumulative decay factor of the previous time step.

[0016] Secondly, the present invention provides an image compressed sensing reconstruction system based on a recursive diffusion model.

[0017] An image compressed sensing reconstruction system based on a recursive diffusion model includes: The image acquisition module is used to acquire the original image to be reconstructed. The sampling module is used to perform block-based compressed sensing sampling on the original image to be reconstructed to obtain complete observation values; The initialization module is used to initialize the complete observations using a pseudo-reverse projection operation to obtain the initial reconstructed image; The reconstruction module is used to perform a complete reconstruction of the initial reconstructed image using a recursive diffusion model. Specifically, in the full image domain, the initial reconstructed image is used as the image estimate for the current iteration. Through the lightweight recursive UNet submodule and the operator conditional loop prior submodule in the recursive diffusion model, multi-step iterative reconstruction is performed to output the final reconstructed image.

[0018] Thirdly, the present invention also provides an electronic device, comprising: a memory for storing executable instructions; and a processor for implementing the above-described image compressed sensing reconstruction method based on a recursive diffusion model when executing the executable instructions stored in the memory.

[0019] Fourthly, the present invention also provides a computer-readable storage medium storing executable instructions for causing a processor to execute the executable instructions to implement the above-described image compressed sensing reconstruction method based on a recursive diffusion model.

[0020] Fifthly, the present invention also provides a computer program product comprising executable instructions stored in a computer-readable storage medium; wherein, when the processor of an electronic device reads the executable instructions from the computer-readable storage medium and executes the executable instructions, the above-described image compressed sensing reconstruction method based on a recursive diffusion model is implemented.

[0021] The above one or more technical solutions have the following beneficial effects: 1. To address the issues of computational redundancy and large number of parameters in existing diffusion models for compressed sensing reconstruction, this invention proposes an image compressed sensing reconstruction method and system based on a recursive diffusion model. By introducing a recursive refinement mechanism and step-by-step memory prior, high-quality image reconstruction is achieved with a small number of iterations, significantly reducing computational and storage overhead. This reduces the computational complexity and number of parameters while maintaining the high reconstruction quality of the diffusion model, effectively improving the model's practicality and generalization ability. This method outperforms existing mainstream methods on multiple datasets and has high practical value and promising prospects for widespread application.

[0022] 2. This invention maintains step memory through the Operator Conditional Cyclic Prior (OCRP) submodule, ensuring consistency and stability during multi-step iterations. Simultaneously, the introduction of physically guided features integrates the physical constraints of compressed sensing, enhancing the model's generalization ability to various image types and complex scenes. By introducing recursive residual refinement units (RMs) with shared weights into the lightweight recursive UNet submodule, it avoids the design of T independent U-Nets corresponding to T iterations in traditional diffusion models. While ensuring reconstruction quality, it significantly reduces the number of model parameters and computational redundancy, lowers storage overhead, and is more suitable for resource-constrained scenarios. Through joint optimization using multiple loss functions—pixel-level loss, multi-scale loss, measurement consistency loss, and projection consistency loss—it ensures that the reconstructed image achieves excellent levels in pixel detail, scale adaptability, and physical constraint satisfaction. The recursive diffusion model designed in this invention integrates the physical constraints of compressed sensing (such as pseudo-inverse projection and data consistency projection), improving the model's interpretability. Furthermore, the lightweight network structure and small number of iterations ensure the model's real-time performance and practicality, making it widely applicable to real-world scenarios such as medical imaging and remote sensing monitoring.

[0023] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0024] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0025] Figure 1 This is a diagram illustrating the overall framework of the image compressed sensing reconstruction method based on a recursive diffusion model proposed in this invention. Figure 2 This is a schematic diagram of a single-step reconstruction process in an embodiment of the present invention; Figure 3 This is a schematic diagram of the lightweight recursive UNet module in an embodiment of the present invention; Figure 4 The images show a comparison of the reconstruction results of the method described in this embodiment of the invention with other existing methods on different datasets; where (a) is the comparison result on the existing Set11 dataset, (b) is the comparison result on the existing BSDS500 dataset, and (c) is the comparison result on the existing Urban100 dataset. Detailed Implementation

[0026] It should be noted that the following detailed descriptions are exemplary and are intended only to describe specific embodiments and to provide further explanation of the invention, and are not intended to limit the scope of exemplary embodiments of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0027] Example 1 To address the computational redundancy and large parameter count issues inherent in existing diffusion models for compressed sensing reconstruction, this embodiment proposes an image compressed sensing reconstruction method based on a recursive diffusion model. This method involves performing block sampling during the compressed sensing sampling stage, reconstructing the original RGB image according to... The image is divided into non-overlapping blocks. After each block is expanded, the observations are sampled from the learned measurement matrix. Then, during the initialization phase, pseudo-inverse backcasting is performed on each block and the image is folded back into the whole image to obtain the initial image estimate for the entire image domain. Then, the reconstruction iteration stage begins, and the reconstruction result is obtained after T iterations, thereby realizing compressed sensing reconstruction processing of images of arbitrary size.

[0028] like Figure 1 As shown, the method proposed in this embodiment specifically includes the following steps: Step S1: Obtain the original image to be reconstructed.

[0029] In this embodiment, the original RGB images from public datasets (such as Set11, BSDS500, Urban100) are selected as the images to be reconstructed, and the image resolution is uniformly adjusted to 256×256.

[0030] Step S2: Perform block-based compressed sensing sampling on the original image to be reconstructed to obtain complete observation values.

[0031] Specifically, first, the original image is divided into... The blocks are divided in a non-overlapping manner. After each block is expanded, it is sampled using the learned measurement matrix (i.e., the compressed sensing sampling matrix, which is obtained through training) to obtain the observation values ​​for each block, as follows: (1) In the above formula, This represents the first segment after the original image is divided into blocks. piece, Indicates to The sampling matrix used during sampling Yes The corresponding observations obtained after sampling; Secondly, the observations of all blocks are combined to obtain the complete observations. .

[0032] Step S3: Initialize the complete observations using a pseudo-reverse projection operation to obtain the initial reconstructed image.

[0033] Specifically, after compressed sensing block sampling of the original image, to facilitate subsequent reconstruction, pseudo-inverse backprojection is performed on each block during the initialization phase, and the image is folded back into the whole image to obtain an initial image estimate for the entire image domain. ,for: (2) In the above formula, Indicates will The complete observations after being combined The pseudo-inverse of the sampling matrix. It is the initial image estimate for the entire map domain, i.e., the initial reconstructed image.

[0034] Step S4: Using a recursive diffusion model, the initial reconstructed image is fully reconstructed. After T iterations, the final reconstructed image is obtained. Specifically, in the full image domain, the initial reconstructed image is used as the image estimate for the current iteration. Multi-step iterative reconstruction is performed using the lightweight recursive UNet submodule and the operator conditional loop prior submodule in the recursive diffusion model, outputting the final reconstructed image.

[0035] Specifically, after obtaining the initial image estimate of the entire graph domain... After (hereinafter referred to as the initial estimation), the image reconstruction stage begins, which iterates across the entire image domain. The process of reconstructing the image in multiple iterative steps is as follows: Figure 2 As shown, it includes: Step S4.1: Based on the image estimation of the current iteration, and combined with the complete observations, construct the physical guidance features. .

[0036] After entering the image reconstruction stage, the complete observations are first combined in each iteration (or each iteration step). Image estimation from the current iteration Medium-structure physical guidance features Specifically: (3) (4) (5) in, It is a compressed sensing sampling matrix; It is the currently predicted image, i.e., the image estimate for the current iteration t; For complete observations; yes and The residual; It is the pseudo-inverse of the compressed sensing sampling matrix. It is the feature correction amount.

[0037] Step S4.2: Input the image estimation and physical guidance features of the current iteration into the operator conditional loop prior submodule to generate the prior image and adaptive fusion gate of the current iteration.

[0038] In obtaining guiding characteristics Subsequently, an operator conditional recurrent prior (OCRP, ConvGRU) submodule based on several convolutional modules and ReLU activation function modules is constructed to guide the features. With the current image The images are input together into the OCRP submodule to maintain step memory, thereby simultaneously producing the prior image for the current step (i.e., the current iteration / step) and the adaptive fusion gate. .

[0039] Step S4.3: Input the image estimate of the current iteration into the lightweight recursive UNet submodule, extract the intermediate features of the current image estimate, and then perform multiple rounds of residual refinement through the recursive unit with shared weights to obtain refined features. Restore the refined features to obtain noise prediction, and perform analytical denoising on the predicted noise to obtain the current image estimate after noise removal.

[0040] Specifically, the recursive UNet submodule (RNR-UNet) runs in parallel with the OCRP submodule, estimating the image in the current iteration. Input into the RNR-UNet submodule, such as Figure 3 As shown, the input image (size H×W, where H is height and W is width) is first converted from 3 channels to 12 channels through pixel shuffle, then fed into the four encoders of UNet to gradually reduce the resolution to expand the receptive field. Intermediate layers based on residual modules fuse cross-layer information and model the global context. Finally, the decoder restores the resolution and outputs the corresponding features. That is, in this submodule, the current reconstructed image is first processed in pixel downsampling space. (i.e., the current image estimate) is dimensionally integrated and encoded to obtain a feature representation with higher channels and lower resolution. Then, an encoder-decoder and a UNet backbone network with cross-layer jump connections are used in conjunction with residual blocks to complete a basic reconstruction and output the intermediate features of the downsampled domain (i.e., the pixel downsampled space).

[0041] Furthermore, to improve estimation accuracy, this module introduces a recursive unit (RM) with shared weights to perform multiple rounds of residual refinement on the basic result (i.e., the intermediate features output above). In this embodiment, the obtained intermediate features are used as the initial noise prediction input to the recursive block (i.e., the recursive unit RM) for recursive noise refinement. The recursive block consists of depthwise convolution, pointwise convolution, Gaussian error, pointwise convolution, and self-attention modules arranged sequentially. The initial noise estimate (i.e., the initial noise prediction) is used as the feature of the initial round, and each round recursively uses the features from the previous round. As input, a residual enhancement amount is generated. The current wheel feature is formed through the residual loop. This approach allows for the gradual injection of richer details while maintaining the same spatial size and number of channels. Furthermore, because recursive units share weights across rounds, the entire process is equivalent to unfolding a multi-layer residual network in the feature space, significantly reducing parameter and memory overhead. After completing... After recursively applying the wheel residuals, the final refined features are obtained. .

[0042] Then, a 3-channel recovery operation is performed by pixel shuffling, that is, the refined features are restored back to the original pixel space through PixelShuffle×2 to obtain the noise prediction. The specific steps for recursively refining the noise features are as follows: (6) In the above formula, This is a characteristic of the previous round of recursion. It is the residual enhancement amount. This is the current recursive characteristic.

[0043] Finally, after obtaining the noise predicted by RNR-UNet, the predicted noise is... Perform analytical denoising as follows: (7) In the above formula, yes t Noisy images of the steps, It is noise predicted by RNR-UNet. It is the cumulative attenuation factor in noise scheduling. It is the current image estimate after noise removal in the current round.

[0044] Step S4.4: Based on the adaptive fusion gate, the prior image and the current image estimate after denoising are adaptively fused, and the data consistency projection alignment measurement constraint is executed. Then, the image estimate for the next iteration is obtained by updating through DDIM (Denoising Diffusion Implicit Models).

[0045] Specifically, after the above steps, in the adaptive fusion gate Under the control of [the system / mechanism], the denoising result is adaptively fused with the prior image, followed by data consistency projection to strictly align with measurement constraints. After consistency projection, DDIM update is performed to obtain the image estimate for the next time step (i.e., the next iteration). The DDIM update formula is as follows: (8) In the above formula, It is noise predicted by RNR-UNet. It is the image estimate after removing noise at the current time. This is the sampling result at the next time step. It is the cumulative decay factor of the previous time step.

[0046] Step S4.5: Continuously iterate until the set number of iterations is reached, and output the final image estimate as the final reconstructed image.

[0047] Through the aforementioned iterative process, OCRP provides stride consistency and stable convergence, RNR-UNet enhances the detail within a single step, data consistency ensures physical constraints at the observation end, and step size scheduling manages the global evolution rhythm. These three elements work together to gradually recover high-quality reconstruction results even with small step sizes, ultimately achieving… At all times, we obtain image outputs with higher consistency with the true value distribution and more complete texture. .

[0048] As one implementation method, in order to obtain high-quality reconstructed images, the loss function for multi-step iterative reconstruction using a recursive diffusion model consists of the following four weighted terms: The pixel-level mean square error loss is: (9) The multi-scale mean square error loss is: (10) The measurement consistency loss is: (11) The projection consistency loss is: (12) The total loss is: (13) In the above formula, It predicts the value of the i-th pixel in the image. It is the value of the i-th pixel in the real image, and N is the total number of pixels in the image; and These are the predicted and real images at different scales, respectively, where M is the number of scales; It involves left-multiplying the predicted image by the compressed sensing operator; These are actual measurement results; It is the pseudo-inverse of the compressed sensing operator; For weighted weights.

[0049] Furthermore, to verify the effectiveness of this method, it was compared with existing mainstream compressed sensing reconstruction methods or models such as ISTA-Net+ (Iterative Shrinkage-Thresholding Algorithm Network Plus), CSNet+ (Compressive Sensing Network Plus), OPINE-Net+ (Optimization-Inspired Explicable Deep Network Plus), COAST (Controllable Arbitrary-Sampling Network), and Csformer (Compressive Sensing Transformer) on the existing Set11, BSDS500, and Urban100 datasets, with sampling rates set to 0.10, 0.25, and 0.50, respectively.

[0050] Through the above comparative experiments, we obtained comparison charts of the reconstruction performance of different methods on different datasets at different sampling rates. Specifically, at low sampling rates, such as Rate=0.10, the experimental results of different methods on different datasets are shown below. Figure 4 As shown in Table 1 below, image quality assessment metrics PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity) were introduced to quantify the experimental results.

[0051] Table 1. Comparative experimental results of different reconstruction methods at different sampling rates

[0052] Clearly, the experimental results of this method outperform existing mainstream methods at multiple sampling rates, achieving higher quality image reconstruction. Furthermore, this method also exhibits optimal reconstruction performance at low sampling rates (e.g., Rate=0.10), further validating its ability to reduce the computational complexity and number of parameters in image reconstruction, thus ensuring the final reconstruction effect. The method proposed in this embodiment enhances the model's practicality and generalization ability, meeting the needs of real-world applications.

[0053] Example 2 This embodiment provides an image compressed sensing reconstruction system based on a recursive diffusion model, including: The image acquisition module is used to acquire the original image to be reconstructed. The sampling module is used to perform block-based compressed sensing sampling on the original image to be reconstructed to obtain complete observation values; The initialization module is used to initialize the complete observations using a pseudo-reverse projection operation to obtain the initial reconstructed image; The reconstruction module is used to perform a complete reconstruction of the initial reconstructed image using a recursive diffusion model. Specifically, in the full image domain, the initial reconstructed image is used as the image estimate for the current iteration. Through the lightweight recursive UNet submodule and the operator conditional loop prior submodule in the recursive diffusion model, multi-step iterative reconstruction is performed to output the final reconstructed image.

[0054] Example 3 This embodiment provides an electronic device, including: a memory for storing executable instructions; and a processor for executing the executable instructions stored in the memory to implement the method provided in this embodiment.

[0055] Example 4 This embodiment also provides a computer-readable storage medium storing executable instructions, which, when executed by a processor, will cause the processor to execute the method described above in this embodiment.

[0056] Example 5 This embodiment provides a computer program product including executable instructions, which are computer instructions; the executable instructions are stored in a computer-readable storage medium. When the processor of an electronic device reads the executable instructions from the computer-readable storage medium and executes the executable instructions, the electronic device performs the method described in this embodiment.

[0057] The steps and methods involved in Embodiments 2 to 5 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0058] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0059] The above description is only a preferred embodiment of the present invention. Although the specific implementation of the present invention has been described in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the present invention.

Claims

1. An image compressed sensing reconstruction method based on a recursive diffusion model, characterized in that, include: Obtain the original image to be reconstructed; The original image to be reconstructed is subjected to block-based compressed sensing sampling to obtain complete observations; A pseudo-reverse projection operation is used to initialize the complete observations to obtain the initial reconstructed image; The recursive diffusion model is used to fully reconstruct the initial image. In the full image domain, the initial reconstructed image is used as the image estimate for the current iteration. Multi-step iterative reconstruction is performed through the lightweight recursive UNet submodule and the operator conditional loop prior submodule in the recursive diffusion model to output the final reconstructed image.

2. The image compressed sensing reconstruction method based on the recursive diffusion model as described in claim 1, characterized in that, The original image to be reconstructed is subjected to block-based compressed sensing sampling to obtain complete observations, including: Press the original image The blocks are divided in a non-overlapping manner. After each block is expanded, it is sampled using the learned measurement matrix to obtain the observation values ​​for each block, as follows: ； In the above formula, This represents the first segment after the original image is divided into blocks. piece, Indicates to The sampling matrix used during sampling Yes The corresponding observations obtained after sampling; By combining the observations from all blocks, a complete set of observations can be obtained. .

3. The image compressed sensing reconstruction method based on the recursive diffusion model as described in claim 1, characterized in that, A pseudo-reverse projection operation is used to initialize the complete observations, resulting in an initial reconstructed image, including: During the initialization phase, a pseudo-reverse projection is performed on each block and then collapsed back to the entire image to obtain an initial image estimate for the entire map domain. ,for: ； In the above formula, Indicates will The complete observations after being combined The pseudo-inverse of the sampling matrix. It is the initial image estimate for the entire map domain, i.e., the initial reconstructed image.

4. The image compressed sensing reconstruction method based on the recursive diffusion model as described in claim 1, characterized in that, The process of multi-step iterative image reconstruction is as follows: Based on the image estimation of the current iteration, combined with complete observations, physical guidance features are constructed. The image estimation and physical guidance features of the current iteration are input into the operator conditional loop prior submodule to generate the prior image and adaptive fusion gate for the current iteration; The image estimate of the current iteration is input into the lightweight recursive UNet submodule, which extracts the intermediate features of the current image estimate. Then, it is refined through multiple rounds of residual refinement by a recursive unit with shared weights to obtain refined features. The refined features are restored to obtain noise prediction. The predicted noise is then denoised analytically to obtain the current image estimate after noise removal. Based on the adaptive fusion gate, the prior image and the current image estimate after denoising are adaptively fused, and the data consistency projection alignment measurement constraint is executed. Then, the image estimate for the next iteration is obtained by updating through DDIM. The process is iterated continuously until the set number of iterations is reached, and the final image estimate is used as the final reconstructed image and then output.

5. The image compressed sensing reconstruction method based on the recursive diffusion model as described in claim 4, characterized in that, The image estimate for the current iteration is input into the lightweight recursive UNet submodule to obtain the noise-removed current image estimate, including: The current image is estimated in the pixel downsampling space, and then the encoder-decoder and the cross-layer skip connection UNet backbone network are used in conjunction with residual blocks to perform basic reconstruction, outputting the intermediate features of the pixel downsampling space. A recursive unit with shared weights is introduced to perform multi-round residual refinement of intermediate features, as follows: each round recursively applies the features from the previous round. As input, the residual enhancement amount is generated. The current wheel feature is formed through the residual loop. The intermediate features are used as the features of the initial round; after... After recursively applying the wheel residuals, the final refined features are obtained; The refined features are restored to the original pixel space using PixelShuffle×2 to obtain the noise prediction. ; Based on the predicted noise Perform analytical denoising as follows: ； In the above formula, It is a noisy image at step t. It is noise predicted by the lightweight recursive UNet. It is the cumulative attenuation factor in noise scheduling. It is the current image estimate after noise removal in the current round.

6. The image compressed sensing reconstruction method based on the recursive diffusion model as described in claim 1, characterized in that, The loss function for multi-step iterative reconstruction using the recursive diffusion model is a weighted sum of pixel-level mean square error loss, multi-scale mean square error loss, measurement consistency loss, and projection consistency loss.

7. An image compressed sensing reconstruction system based on a recursive diffusion model, characterized in that, include: The image acquisition module is used to acquire the original image to be reconstructed. The sampling module is used to perform block-based compressed sensing sampling on the original image to be reconstructed to obtain complete observation values; The initialization module is used to initialize the complete observations using a pseudo-reverse projection operation to obtain the initial reconstructed image; The reconstruction module is used to perform a complete reconstruction of the initial reconstructed image using a recursive diffusion model. Specifically, in the full image domain, the initial reconstructed image is used as the image estimate for the current iteration. Through the lightweight recursive UNet submodule and the operator conditional loop prior submodule in the recursive diffusion model, multi-step iterative reconstruction is performed to output the final reconstructed image.

8. An electronic device, characterized in that, include: Memory, used to store executable instructions; The processor, when executing executable instructions stored in the memory, implements the image compressed sensing reconstruction method based on the recursive diffusion model as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The device stores executable instructions that, when executed by a processor, implement the image compressed sensing reconstruction method based on a recursive diffusion model as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes executable instructions stored in a computer-readable storage medium; When the processor of the electronic device reads the executable instructions from the computer-readable storage medium and executes the executable instructions, it implements the image compressed sensing reconstruction method based on the recursive diffusion model as described in any one of claims 1-6.

Images