An image compression sensing reconstruction method and system of an uncertainty correction reversible diffusion model

Abstract

The application discloses an image compression sensing reconstruction method and system of an uncertainty correction reversible diffusion model, and belongs to the field of image processing. The method embeds an uncertainty refinement module in a reversible diffusion model framework: at each step of reverse diffusion, the noise is predicted and removed by a noise prediction module to obtain an initial image, which is projected into the null space of the measurement matrix; the null space projection image, the physical measurement residual and the high-frequency edge prior are spliced into a multi-dimensional feature, which is input into the uncertainty refinement module; the module adopts a serial double-attention mechanism of channel attention and spatial attention to locate artifacts, and after a residual block, the uncertainty gate is used to adaptively control the correction amplitude, and the candidate residual is projected into the null space and fused with the reference image; after multiple iterations, the reconstruction result is output. The application introduces a pixel, perception and frequency domain combined loss to suppress spectral bias, which can effectively eliminate artifacts and restore texture at a very low sampling rate.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and specifically to an image compressed sensing reconstruction method and system based on an uncertainty-corrected reversible diffusion model. Background Technology

[0002] Compressed sensing (CS) [1] achieves synchronous sampling and compression of signals through linear random transformation, and can reconstruct the original signal using only a small number of observations. The number of observations required is far lower than that required by the sub-Nyquist sampling rate. For signals Sampling process The measurement vector can be obtained. ,in Given a sampling matrix. The sampling rate is Compressed sensing has spawned a range of imaging applications, including single-pixel imaging (SPI)[2], magnetic resonance imaging (MRI)[3], computed tomography (CT)[4], and snapshot compression imaging (SCI)[5]. However, the inverse problem is ill-posed, reconstructing from finite measurements It is quite difficult and often produces artifacts or overly smooth reconstruction results.

[0003] In the field of compressed sensing (CS) reconstruction of natural images, traditional optimization-based methods [6][7] do not require training and have high interpretability, but suffer from low reconstruction accuracy, poor computational efficiency, and cumbersome parameter tuning. With the emergence of deep neural networks (NNs), end-to-end neural network methods have fast inference speed, can be adapted to dedicated compressed sensing sampling matrices, and have strong controllability, outperforming traditional optimization-based methods. Early end-to-end neural networks [8] treated compressed sensing reconstruction as a dealiasing problem and achieved fast non-iterative reconstruction through a single forward propagation of neural network layers. After that, scholars proposed deep unfolded networks [9]-

[12] to improve the interpretability of neural networks. Recently, end-to-end neural networks based on Transformers

[13]

[14] have been proposed to further improve the model's representation ability and reconstruction performance. However, such methods usually require designing and training a completely new network structure from scratch, which is very time-consuming, and the final performance of the model often does not reach the optimal level.

[0004] The development of compressed sensing reconstruction methods based on diffusion models is very rapid. The powerful generation capabilities based on diffusion models are mainly divided into methods guided by posterior sampling and end-to-end diffusion methods. Methods guided by posterior sampling

[17] mainly utilize the powerful image prior of the pre-trained diffusion model. In the back-diffusion iteration process, the compression sensing measurement consistency constraint is added to guide the sampling process to approach the real image, requiring almost no or only a small amount of retraining for the compression sensing task. For example, DDNM

[18] and PSLD

[19] use the posterior sampling guided method. However, this type of method has many iterations, slow inference speed, and is highly dependent on the quality of the pre-trained diffusion model. Unlike the posterior sampling guided method, the end-to-end diffusion method does not depend on the pre-trained diffusion model, and its performance is more suitable for the compression sensing task. For example, IDM

[20] transforms the sampling step into a reversible architecture, which can perform memory-efficient fine-tuning of large pre-trained models on compressed sensing measurement data. However, under extremely low sampling rate conditions, the diffusion model still has spectral bias and generation artifact problems. Spectral bias will cause the model to focus on low-frequency components, resulting in excessively smooth textures. Meanwhile, compressed sensing often generates false textures in the physical null space that do not conform to the true physical distribution. How to leverage the powerful generative capabilities of diffusion models while ensuring strict physical fidelity remains an important unsolved problem. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an image compressed sensing reconstruction method and system using an uncertainty-corrected reversible diffusion model, which can alleviate the problems of spectral bias and generation artifacts in low-sampling-rate image compressed sensing.

[0006] This invention discloses an image compressed sensing reconstruction method using an uncertainty-corrected reversible diffusion model, comprising the following steps: Step 1: Input the sampling matrix A and the observed compressed sensing measurement value y; Step 2: At the current time step t of the back diffusion process, predict the noise using the noise prediction module. And remove noise to obtain the predicted clean image. The predicted clean image is projected onto the null space of the sampling matrix A to obtain the null-space projection image. ; Step 3: Project the null space image Multidimensional input features are constructed by concatenating physical measurement residuals with high-frequency edge priors. ; Step 4: Transfer the multidimensional input features The input uncertainty refinement module performs adaptive refinement processing and outputs the reconstructed image at the next time step t-1. ; Step 5: Iterate steps 2 to 4 for t-1 times to output the final reconstructed image. .

[0007] As a further improvement of the present invention, in step 2, Noise prediction and null space constraints are performed during the backdiffusion process using a backdiffusion model. At step t of the backdiffusion, the model uses a noise prediction module to predict the noise. And calculate the initial image. : in, This represents the noisy image at the current time step. This represents the noise predicted by the noise prediction module. This represents the cumulative weight of the original image signal after t forward diffusions; , It is a pre-set noise scheduling coefficient; The formula for projecting the predicted clean image onto the null space to obtain the null-space projected image is: in, This represents the observed compressed sensing measurements. It is the observation matrix. The pseudo-inverse of the observation matrix.

[0008] As a further improvement of the present invention, in step 3, the formula for constructing multidimensional input features is: in, Indicates splicing, It is the physical measurement residual gradient, which represents the offset direction between the current image and the actual measurement; Used to extract the initial high-frequency edges of the image.

[0009] As a further improvement of the present invention, in step 4, the uncertainty refinement module adopts a serial dual attention mechanism. First, the channel attention module extracts the inter-channel dependencies. The channel attention module performs global average pooling and max pooling in the spatial dimension. Then, it captures nonlinear correlations with the help of a shared multilayer perceptron. The formula is as follows: in, This represents the Sigmoid activation function. This indicates element-wise multiplication. , , .

[0010] As a further improvement of the present invention, after the channel attention processing is completed, a two-dimensional spatial mask is generated by the spatial attention module to locate compressed perception artifacts and structural edges. The spatial attention module concatenates the channel-wise average pooling feature map and the max pooling feature map, and uses a large receptive field convolution to obtain local context information. The formula is as follows: in, This indicates element-wise multiplication. This represents the Sigmoid activation function. Indicates splicing, Indicates average pooling This indicates max pooling. This indicates a convolution operation.

[0011] As a further improvement of the present invention, the feature F″ after dual attention filtering is processed by two residual blocks to obtain F″′, and the correction magnitude is dynamically controlled by a gating mechanism based on uncertainty estimation: the feature is simultaneously input into two convolution heads to obtain the prediction candidate residual r and the uncertainty parameter s respectively; the Softplus function is used to ensure that the uncertainty parameter is positive, and then the tanh function is used to map it to the (0,1) interval to obtain the gating coefficient g: in, This represents a 1x1 convolutional network used to predict current spatial features. The uncertainty signal is used to characterize the variance of the image. This is used to ensure that the output is positive. This indicates a normalization operation.

[0012] As a further improvement of the present invention, the candidate residual r is projected onto the null space of the measurement matrix A to obtain the null space residual. : in, The 1x1 convolutional network is used to transform deep features to the image domain to obtain texture and edge information of the image, where I is the identity matrix; the final optimized image consists of the baseline prediction result and the gated null space residual. .

[0013] As a further improvement of the present invention, the prior hidden state is derived from the image obtained by the uncertainty refinement module through DDIM sampling, and the state transition equation is: .

[0014] This invention also discloses an image compressed sensing reconstruction system based on an uncertainty-corrected reversible diffusion model, comprising: The noise prediction module is used to predict noise at the current time step of the back diffusion process, remove noise to obtain a clean predicted image and project it onto the null space of the measurement matrix. The feature construction module is used to concatenate the null-space projection image, physical measurement residuals and high-frequency edge priors into multi-dimensional input features; The uncertainty refinement module, including a channel attention unit, a spatial attention unit, a residual block, and an uncertainty gating unit connected in series, is used to adaptively refine the multidimensional input features and output the reconstructed image at the next time step. The iterative control module is used to control the noise prediction module, feature construction module, and uncertainty refinement module to iterate t-1 times and output the final reconstructed image.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention seamlessly embeds a lightweight uncertainty refinement module into a reversible diffusion architecture. It introduces physical null space constraints and uncertainty-guided adaptive refinement at each step of the back diffusion process. While leveraging the powerful generation capabilities of the diffusion model, it ensures strict physical fidelity and effectively suppresses artifacts in the physical null space under extremely low sampling rate conditions.

[0016] 2. This invention employs a serial dual attention mechanism consisting of channel attention and spatial attention. Channel attention assesses the importance of each channel and amplifies the weight of high-frequency and physical constraint information, while spatial attention locates artifacts and structural edges, enabling precise identification of content and location requiring refinement.

[0017] 3. The uncertainty gating mechanism proposed in this invention estimates the confidence level of image restoration, injects high-frequency details in high-confidence regions, and suppresses corrections in smooth regions, thereby reducing the generation of false textures from the source.

[0018] 4. This invention introduces FFT frequency domain loss to impose an explicit penalty on spectral deviation, driving the backpropagation gradient to preferentially recover high-frequency components, effectively alleviating the problem of excessive texture smoothing caused by pixel domain regression loss.

[0019] 5. Extensive experiments on the Set11, CBSD68, DIV2K, and Urban100 datasets demonstrate that this invention achieves competitive objective performance metrics and significantly outperforms current state-of-the-art methods in terms of subjective visual effects and texture fidelity. Attached Figure Description

[0020] Figure 1A compressed sensing framework diagram of the uncertainty correction reversible diffusion model provided in the embodiments of the present invention; Figure 2 A flowchart of a compressed sensing reconstruction method for an uncertainty-corrected reversible diffusion model provided in an embodiment of the present invention; Figure 3 An uncertainty refinement module framework diagram provided for embodiments of the present invention; Figure 4 A comparison of compressed sensing reconstruction results for two images, “house” (top) and “foreman” (bottom), in the IV2K dataset under a compressed sensing sampling rate of γ=10% provided in this embodiment of the invention. Figure 5 Comparison of compressed sensing reconstruction results for IV2K dataset images “0826” and “0861” (top image) and CBSD68 dataset image “test 44” (bottom image) under the condition of compressed sensing sampling rate γ = 10% provided in the embodiments of the present invention; Figure 6 Under the condition of compressed sensing sampling rate γ = 10% provided in the embodiments of the present invention, the results of compressed sensing reconstruction of two images named "img 60" (top image) and "img 46" (bottom image) in the Urban100 dataset are compared. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The present invention will now be described in further detail with reference to the accompanying drawings: Diffusion models have become powerful generative prior models in the field of compressed sensing (CS). However, under extremely low sampling rates, these models suffer from spectral bias (oversmoothing) and artifact illusions in the physical null space. To address these issues, this invention proposes an Uncertainty-Correction Invertible Diffusion Model (UIDM), seamlessly embedding a lightweight uncertainty correction module into a reversible diffusion architecture. This framework employs a serial dual attention mechanism to coordinate channel and spatial features, enabling precise localization of structured artifacts. To solve the pixel-domain oversmoothing problem, this invention introduces a Fast Fourier Transform (FFT) loss function to explicitly penalize spectral bias and restore clear textures. Furthermore, the image restoration confidence is evaluated through an adaptive gating mechanism in the uncertainty correction module, injecting high-frequency details only in high-confidence regions to suppress artifact illusions at their source. Extensive experiments conducted on the Set11, CBSD68, DIV2K, and Urban100 datasets demonstrate that UIDM achieves competitive objective performance metrics and significantly outperforms current state-of-the-art methods in terms of subjective visual effects and texture fidelity.

[0023] This invention provides an image compressed sensing reconstruction method and system based on an uncertainty-corrected reversible diffusion model. It constructs an embedded lightweight uncertainty correction module based on the efficient reversible diffusion model (IDM)

[20] framework, and applies the input observation matrix... And the observed compressed sensing measurements Reconstruction is required. The model framework is as follows: Figure 1 As shown, its network framework mainly includes noise prediction and uncertainty refinement modules. Its process is as follows: Figure 2 As shown, first, input the observation matrix. And the observed compressed sensing measurements Then, the model obtains the noise estimation U-Net module. Then, noise is removed to obtain the predicted clean image. The predicted clean image is projected onto the physical null space, thus providing the model with explicit error correction prior information. The image projected onto the physical null space, the physical measurement residuals, and the high-frequency edge priors are then concatenated to construct a multi-dimensional feature input uncertainty refinement module (URM) for processing, adaptively eliminating the aforementioned artifacts. This iterative process is repeated t-1 times, and the final optimized reconstruction result is output. .

[0024] This invention discloses an image compressed sensing reconstruction method using an uncertainty-corrected reversible diffusion model, comprising the following steps: Step 1: Input the sampling matrix A and the observed compressed sensing measurement value y; Step 2: At the current time step t of the back diffusion process, predict the noise using the noise prediction module. And remove noise to obtain the predicted clean image. The predicted clean image is projected onto the null space of the sampling matrix A to obtain the null-space projection image. ; Step 3: Project the null space image Multidimensional input features are constructed by concatenating physical measurement residuals with high-frequency edge priors. ; Step 4: Transfer the multidimensional input features The input uncertainty refinement module performs adaptive refinement processing and outputs the reconstructed image at the next time step t-1. ; Step 5: Iterate steps 2 to 4 for t-1 times to output the final reconstructed image. .

[0025] The main contents of the noise prediction module in steps 2 and 3 above are described below: Noise prediction and null space constraints are performed during the backdiffusion process using the inverse diffusion model (IDM)

[20] . In the t-th step of the backdiffusion, the model uses the noise prediction module to predict the noise. And calculate a rough initial image. : (1) in, This represents the noisy image at the current time step. This represents the noise predicted by the noise prediction module. This represents the cumulative weight of the original image signal after t forward diffusions; , It is a pre-set noise control factor.

[0026] To ensure that the reconstruction results meet the physical observation requirements Typically, a measurement matrix is ​​used. Applying a data consistency projection to it will Projecting onto the null space yields the baseline result. : (2) in, This represents the observed compressed sensing measurements. It is the observation matrix. The pseudo-inverse of the observation matrix.

[0027] Although physical projection pulls the image back into the data manifold, it is limited by The prediction error The internal structure often retains aliasing artifacts and severely lacks high-frequency textures. Therefore, to provide sufficient error correction cues for the uncertainty refinement module, this invention concatenates the image projected onto the physical null space, physical measurement errors, and high-frequency edge priors to construct a multi-dimensional input feature. : (3) in, Indicates splicing, It is the physical measurement residual gradient, which represents the offset direction between the current image and the actual measurement; Used to extract the initial high-frequency edges of the image.

[0028] Finally, multidimensional features It is then sent to the uncertainty refinement module for adaptive fine-tuning.

[0029] In step 4, as Figure 3 As shown, the main contents of the uncertainty refinement module are described below: Traditional residual inpainting networks tend to apply the same intensity of correction across the entire image, which can easily introduce generational illusions that do not conform to the true physical distribution in the large null space of compressed sensing tasks. To address this, this invention integrates a dual attention mechanism and uncertainty gating into the uncertainty refinement module. This design enables the network not only to identify "content that needs refinement" and "location that needs refinement," but also to intelligently determine "whether refinement is necessary," thereby achieving a reasonable balance between generative richness and physical fidelity.

[0030] Considering the significant differences in information density of input features across different channels and spatial locations, this invention introduces a serial dual attention mechanism consisting of channel attention and spatial attention.

[0031] To account for the starkly different sparsity characteristics between physical residuals and images, this invention employs a Channel Attention (CA) module to extract inter-channel dependencies and assess the importance of each channel. This module first performs global average pooling and max pooling in the spatial dimension, and then leverages a shared multilayer perceptron (MLP) to capture non-linear correlations. (4) in, This represents the Sigmoid activation function. This indicates element-wise multiplication. , , This step dynamically amplifies the channel weights of high-frequency and physical constraint information.

[0032] After channel reweighting, the Spatial Attention (SA) module generates a two-dimensional spatial mask to locate compressed perceptual artifacts and structural edges. This module concatenates the channel-wise average pooling feature maps and max pooling feature maps, and then uses a large receptive field convolution to obtain a wide range of local contextual information. (5) in, This indicates element-wise multiplication. This represents the Sigmoid activation function. Indicates splicing, Indicates average pooling This indicates max pooling. This indicates a convolution operation.

[0033] Features filtered by dual attention The response in the high-frequency artifact region is significantly enhanced, while smooth background noise is effectively suppressed.

[0034] This invention inputs the filtered feature F′′ into two simple residual blocks (Res) to obtain F′′′: (6) Subsequently, this invention proposes a gating mechanism based on uncertainty estimation to dynamically control the correction magnitude. Simultaneously inputting two different convolution heads yields prediction candidate residuals containing high-frequency information. and uncertainty parameters This uncertainty-guided gating mechanism can dynamically adjust the refinement intensity based on the network's prediction confidence level.

[0035] To construct the gated signal, this invention first uses the Softplus function to ensure that the uncertainty parameter remains positive, and then uses the tanh function to map it to the open interval (0, 1) to obtain the gate coefficient. : (7) (8) Specifically, This represents a 1x1 convolutional network used to predict current spatial features. The uncertainty signal is used to characterize the variance of the image; Used to ensure that the output is positive (actually performs the function of absolute value operation). This indicates a normalization operation.

[0036] Gating coefficient It is a modulation factor that senses confidence. In regions where the network can identify artifacts with high confidence, It asymptotically approaches 1, thus achieving precise correction. Conversely, in blurred or smooth areas, By approaching 0, potential artifacts are suppressed, achieving a conservative reconstruction that conforms to physical laws.

[0037] Meanwhile, to ensure that the image after adding residuals strictly meets the physical measurement constraints, this invention uses candidate residuals... Projected onto the measurement matrix The null space is obtained, and the null space residual is obtained. : (9) (10) Specifically, This represents a 1x1 convolutional network used to transform deep features into the image domain to obtain texture and edge information. The final optimized image consists of the baseline prediction and the gated null space residual. (11) This mechanism ensures that the generative model behaves very cautiously when adding high-frequency details, essentially eliminating the generation of false textures. The image obtained after optimization by the uncertainty refinement module. The input state transition equation is then sampled using DDIM to derive the prior hidden states. : (12) Loss Strategy To suppress frequency domain smoothing caused by regression loss during end-to-end training of the diffusion model, this invention designs a combined loss function jointly driven by the pixel domain, perceptual domain, and frequency domain. : (13) in, (1) Pixel loss This invention adopts Loss to constrain reconstruction results With real images Absolute consistency in pixel space: (14) in, This represents the ground truth of the image.

[0038] (2) Perceived loss To make the reconstructed image more consistent with the perceptual preferences of the human visual system, this invention introduces a pre-trained VGG network to extract features and calculate the LPIPS perceptual distance. This effectively alleviates the... Excessive ambiguity caused by loss: (15) in, This represents the feature extraction process of the VGG network.

[0039] (3) Frequency domain loss To compensate for the inherent spectral bias in the diffusion process, this invention introduces a frequency domain constraint: a two-dimensional Fast Fourier Transform (FFT) is used to penalize the amplitude spectrum difference between the reconstructed image and the real image. (16) in, This represents a two-dimensional Fast Fourier Transform. Since compressed sensing sampling causes severe loss of high-frequency signals, direct supervision in the frequency domain is used to drive backpropagation gradients to preferentially recover high-frequency components.

[0040] Example: 1. Dataset The network framework of this invention is trained on the WED dataset. The WED dataset includes 4744 images, and the test set consists of test images from the Set11, CBSD68, DIV2K, and Urban100 datasets.

[0041] 2. Comparison Algorithm This invention compares the proposed method with ReconNet [8], ISTA-Net+

[15] , CSNet+

[21] , SCSNet

[22] , OPINE-Net+ [9], DPA-Net

[10] , MAC-Net

[11] , RK-CCSNet

[12] , ISTA-Net++

[23] , COAST

[24] , AMP-Net

[16] , MADUN

[25] , DGUNet+

[26] , MR-CCSNet+

[27] , FSOINet

[28] , CASNet

[29] , TransCS

[14] , OCTUF+

[30] , CSformer

[13] , PRL-PGD+

[31] , IDM

[20] on the Set11, CBSD68, DIV2K, and Urban100 datasets. All deep learning-based comparison methods are tested using their pre-trained models.

[0042] 3. Experimental Setup In the experiment, the image size of the dataset of this invention was adjusted to... Using a block size of Structure of random orthogonal independent and identically distributed Gaussian matrices The framework of this invention is implemented using PyTorch on an NVIDIA RTX 5880 GPU. The batch size is 8. The initial learning rate is set to... The attenuation is achieved using a cosine annealing strategy. In the combined loss strategy, the weight of the pixel-level L1 loss is set to 1.0, and the regularization coefficients of the perceptual loss and the FFT frequency domain loss are strictly set to... and All experiments were trained using the AdamW optimizer.

[0043] Table 1-21 shows the average objective results of the comparison methods on the Set11, CBSD68, DIV2K, and Urban100 datasets. Table 2-8 shows the average subjective results of the comparison methods on the Set11, CBSD68, DIV2K, and Urban100 datasets. The advantages of this invention are: To verify the effectiveness of the proposed compressed sensing method based on the uncertainty correction reversible diffusion model, this invention was tested on datasets in various scenarios to verify the adaptability and effectiveness of the proposed method, including the Set11, CBSD68, DIV2K, and Urban100 datasets. This invention compares the model with several state-of-the-art methods, including ReconNet [8], ISTA-Net+

[15] , CSNet+

[21] , SCSNet, OPINE-Net+ [9], DPA-Net

[10] , MAC-Net

[11] , RK-CCSNet

[12] , ISTA-Net++

[23] , COAST

[24] , AMP-Net

[16] , MADUN

[25] , DGUNet+

[26] , MR-CCSNet+

[27] , FSOINet

[28] , CASNet

[29] , TransCS

[14] , OCTUF+

[30] , CSformer

[13] , PRL-PGD+

[31] , and IDM

[20] . This invention uses PSNR, SSIM, FID, and LPIPS to evaluate image quality.

[0044] Table 1 shows the average augmentation performance of 21 competing methods on the Set11, CBSD68, DIV2K, and Urban100 datasets. It can be seen that the model of this invention is highly competitive in objective metrics (PSNR, SSIM). Meanwhile, Table 2 shows a significant improvement in subjective metrics (FID, LPIPS). Figure 4 , Figure 5 and Figure 6 The visualizations show the results of testing on different datasets. As can be seen, the model of this invention has significant improvements in detail and provides optimal subjective performance.

[0045] References [1]. L. Donoho, “Compressed Sensing,” IEEE Transactions on Information Theory, 2006. [2]. 2. M. F. Duarte, M. A. Davenport, D. Takhar, J. N. Laska, T.Sun, K. F. Kelly, and R. G. Baraniuk, “Single-Pixel Imaging via CompressiveSampling,” IEEE Signal Processing Magazine, 2008. [3]. M. Lustig, D. L. Donoho, and J. M. Pauly, “Sparse MRI: TheApplication of Compressed Sensing for Rapid MR Imaging,”Magnetic Resonance inMedicine,2007. [4]. G.-H. Chen, J. Tang, and S. Leng, “Prior Image ConstrainedCompressed Sensing(PICCS): A Method to Accurately Reconstruct Dynamic CTImages from Highly Undersampled Projection Data Sets,” Medical Physics, 2008. [5]. J. Suo, W. Zhang, J. Gong, X. Yuan, D. J. Brady, and Q. Dai,“Computational Imaging and Artificial Intelligence: The Next Revolution ofMobile Vision,” Proceedings of the IEEE,2023. [6]. J. Zhang, D. Zhao, and W. Gao, “Group-Based SparseRepresentation for Image Restoration,” IEEE Transactions on Image Processing,vol. 23, no. 8, pp.3336–3351, 2014. [7]. W. Dong, G. Shi, X. Li, Y. Ma, and F. Huang, “CompressiveSensing via Nonlocal Low-Rank Regularization,” IEEE Transactions on ImageProcessing, vol. 23, no8, pp. 3618–3632, 2014. [8]. K. Kulkarni, S. Lohit, P. Turaga, R. Kerviche, and A. Ashok,“ReconNet: Noniterative Reconstruction of Images from Compressively SensedMeasurements,” in(CVPR), 2016. [9]. J. Zhang, C. Zhao, and W. Gao, “Optimization-Inspired CompactDeep Compressive Sensing,” IEEE Journal of Selected Topics in SignalProcessing, 2020.

[10] . Y. Sun, J. Chen, Q. Liu, B. Liu, and G. Guo, “Dual-PathAttention Network for Compressed Sensing Image Reconstruction,” IEEETransactions on Image Processing, 2020.

[11] . J. Chen, Y. Sun, Q. Liu, and R. Huang, “Learning MemoryAugmented Cascading Network for Compressed Sensing of Images,”in (ECCV),2020.

[12] . R. Zheng, Y. Zhang, D. Huang, and Q. Chen, “SequentialConvolution and Runge-Kutta Residual Architecture for Image CompressedSensing,”in (ECCV), 2020.

[13] . D. Ye, Z. Ni, H. Wang, J. Zhang, S. Wang, and S. Kwong,“CSformer: Bridging Convolution and Transformer for Compressive Sensing,”IEEETransactions on Image Processing, 2023.

[14] . M. Shen, H. Gan, C. Ning, Y. Hua, and T. Zhang, “Transcs:Atransformer-based hybrid architecture for image compressed sensing,” IEEETransactions on Image Processing, 2022.

[15] . J. Zhang and B. Ghanem, “ISTA-Net: Interpretable Optimization-Inspired Deep Network for Image Compressive Sensing,”in (CVPR), 2018.

[16] . Z. Zhang, Y. Liu, J. Liu, F. Wen, and C. Zhu, “AMP-Net:Denoising-based Deep Unfolding for Compressive Image Sensing,”IEEETransactions on Image Processing, 2021.

[17] . H. Chung, J. Kim, M. T. Mccann, M. L. Klasky, and J. C. Ye,“Diffusion Posterior Sampling for General Noisy Inverse Problems,” in(ICLR),2023.

[18] . Y. Wang, J. Yu, and J. Zhang, “Zero-shot image restorationusing denoising diffusion null-space model,”in (ICLR), 2023.

[19] . L. Rout, N. Raoof, G. Daras, C. Caramanis, A. G. Dimakis, andS. Shakkottai,“Solving Linear Inverse Problems Provably via PosteriorSampling with Latent Diffusion Models,” in (NeurIPS), 2023.

[20] . B. Chen, Z. Zhang, W. Li, C. Zhao, J. Yu, S. Zhao, J. Chen, andJ. Zhang, “Invertible Diffusion Models for Compressed Sensing,”in (CVPR)2025.

[21] . W. Shi, F. Jiang, S. Liu, and D. Zhao, “Image CompressedSensing Using Convo-lutional Neural etwork,”IEEE Transactions on ImageProcessing, 2019.

[22] . W. Shi, F. Jiang, S. Liu, and D. Zhao, “Scalable ConvolutionalNeural Network for Image Compressed Sensing,”in (CVPR), 2019.

[23] . D. You, J. Xie, and J. Zhang, “ISTA-Net++: Flexible DeepUnfolding Network for Compressive Sensing,”in (ICME).

[24] . D. You, J. Zhang, J. Xie, B. Chen, and S. Ma, “COAST:COntrollable Arbitrary-Sampling NeTwork for Compressive Sensing,”IEEETransactions on Image Processing, 2021.

[25] . J. Song, B. Chen, and J. Zhang, “Memory-Augmented DeepUnfolding Network for Compressive Sensing,” in (ACM MM), 2021.

[26] . C. Mou, Q. Wang, and J. Zhang, “Deep generalized unfoldingnetworks for image restoration,” in (CVPR), 2022.

[27] . Z.-E. Fan, F. Lian, and J.-N. Quan, “Global Sensing andMeasurements Reuse for Image Compressed Sensing,” in (CVPR), 2022.

[28] . . W. Chen, C. Yang, and X. Yang, “FSOINet: Feature-SpaceOptimization-Inspired Network for Image Compressive Sensing,” in (ICASSP),2022.

[29] . B. Chen and J. Zhang, “Content-aware Scalable Deep CompressedSensing,” IEEE Transactions on Image Processing, 2022.

[30] . J. Song, C. Mou, S. Wang, S. Ma, and J. Zhang, “Optimizationinspired cross-attention transformer for compressive sensing,” in (CVPR), 2023.

[31] . B. Chen, J. Song, J. Xie, and J. Zhang, “Deep Physics-GuidedUnrolling Generalization for Compressed Sensing,” International Journal of Computer Vision, 2023 The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An image compressed sensing reconstruction method using an uncertainty-corrected reversible diffusion model, characterized in that, Includes the following steps: Step 1: Input the sampling matrix A and the observed compressed sensing measurement value y; Step 2: At the current time step t of the back diffusion process, predict the noise using the noise prediction module. And remove noise to obtain the predicted clean image. The predicted clean image is projected onto the null space of the sampling matrix A to obtain the null-space projection image. ; Step 3: Project the null space image Multidimensional input features are constructed by concatenating physical measurement residuals with high-frequency edge priors. ; Step 4: Transfer the multidimensional input features The input uncertainty refinement module performs adaptive refinement processing and outputs the reconstructed image at the next time step t-1. ; Step 5: Iterate steps 2 to 4 for t-1 times to output the final reconstructed image. .

2. The method according to claim 1, characterized in that, In step 2, Noise prediction and null space constraints are performed during the backdiffusion process using a backdiffusion model. At step t of the backdiffusion, the model uses a noise prediction module to predict the noise. And calculate the initial image. : in, This represents the noisy image at the current time step. This represents the noise predicted by the noise prediction module. This represents the cumulative weight of the original image signal after t forward diffusions; , It is a pre-set noise scheduling coefficient; The formula for projecting the predicted clean image onto the null space to obtain the null-space projected image is: in, This represents the observed compressed sensing measurements. It is the observation matrix. The pseudo-inverse of the observation matrix.

3. The method according to claim 1, characterized in that, In step 3, the formula for constructing multidimensional input features is: in, Indicates splicing, It is the physical measurement residual gradient, which represents the offset direction between the current image and the actual measurement; Used to extract the initial high-frequency edges of the image.

4. The method according to claim 1, characterized in that, In step 4, the uncertainty refinement module employs a serial dual attention mechanism. First, it extracts the inter-channel dependencies through a channel attention module, which performs global average pooling and max pooling in the spatial dimension. Then, it captures nonlinear associations using a shared multilayer perceptron, as shown in the formula: in, This represents the Sigmoid activation function. This indicates element-wise multiplication. , , .

5. The method according to claim 4, characterized in that, After channel attention processing, a two-dimensional spatial mask is generated by the spatial attention module to locate compressed perception artifacts and structural edges. The spatial attention module concatenates the channel-wise average pooling feature map and the max pooling feature map, and uses a large receptive field convolution to obtain local contextual information. The formula is as follows: in, This indicates element-wise multiplication. This represents the Sigmoid activation function. Indicates splicing, Indicates average pooling This indicates max pooling. This indicates a convolution operation.

6. The method according to claim 5, characterized in that, The feature F″ after dual attention filtering is processed by two residual blocks to obtain F″′: 。 7. The method according to claim 6, characterized in that, The correction magnitude is dynamically controlled through a gating mechanism based on uncertainty estimation: features are simultaneously input into two convolutional heads to obtain the prediction candidate residual r and uncertainty parameter s, respectively; the Softplus function is used to ensure that the uncertainty parameter is positive, and then the tanh function is used to map it to the (0,1) interval to obtain the gating coefficient g. in, This represents a 1x1 convolutional network used to predict current spatial features. The uncertainty signal is used to characterize the variance of the image. This is used to ensure that the output is positive. This indicates a normalization operation. Projecting the candidate residual r onto the null space of the measurement matrix A yields the null space residual. : in, This represents a 1x1 convolutional network used to transform deep features into the image domain to obtain texture and edge information of the image; I is the identity matrix. The final optimized image consists of the baseline prediction result and the gated null space residual: 。 8. The method according to claim 1, characterized in that, The image optimized by the uncertainty refinement module is used to derive the prior hidden state through DDIM sampling. The state transition equation is: 。 9. An image compressed sensing reconstruction system with an uncertainty-corrected reversible diffusion model, used to implement the method as described in any one of claims 1 to 8, characterized in that, include: The noise prediction module is used to predict noise at the current time step of the back diffusion process, remove noise to obtain a clean predicted image and project it onto the null space of the measurement matrix. The feature construction module is used to concatenate the null-space projection image, physical measurement residuals and high-frequency edge priors into multi-dimensional input features; The uncertainty refinement module, including a channel attention unit, a spatial attention unit, a residual block, and an uncertainty gating unit connected in series, is used to adaptively refine the multidimensional input features and output the reconstructed image at the next time step. The iterative control module is used to control the noise prediction module, feature construction module, and uncertainty refinement module to iterate t-1 times and output the final reconstructed image.

Images