An image compressive sensing reconstruction method based on a super-gradient denoising network

Abstract

This invention relates to the field of image processing, specifically to an image compressed sensing reconstruction method based on a hypergradient denoising network. Building upon traditional block-by-block image sampling and initial reconstruction, this invention maps the compressed sensing iterative denoising reconstruction model into an end-to-end denoising network using the concept of depth unfolding. Residual dense blocks are introduced into the denoising network as denoising operators to improve the quality of image denoising and reconstruction. Furthermore, a hypergradient strategy is used to optimize the iterative denoising model, accelerating network convergence. This invention overcomes the limitations of manually setting regularization terms by learning noise information through denoising operators during the iterative denoising process, flexibly representing the prior information of the image, and accurately and efficiently reconstructing the original image.

Description

Technical Field

[0001] This invention relates to the field of image processing, and mainly to an image compressed sensing reconstruction method based on a super gradient denoising network. Background Technology

[0002] The emerging theory of compressed sensing leverages the sparse representation of signals to accurately reconstruct the original signal from far fewer measurements than suggested by the Nyquist sampling theorem using nonlinear reconstruction algorithms. Compressed sensing synchronizes sampling and compression, significantly reducing the hardware overhead of signal sampling, transmission, and storage. By shifting the computational burden to the reconstruction process, compressed sensing offers a solution for the acquisition and transmission of digital images in the era of big data. In compressed sensing research, how to reconstruct the original high-dimensional signal from noisy low-dimensional linear observations is a topic of widespread interest.

[0003] Current compressed sensing image reconstruction methods can be mainly divided into two categories. The first is model-based iterative reconstruction algorithms. These algorithms typically utilize the sparsity of images in specific domains such as the discrete cosine transform domain and wavelet domain to establish constrained optimization models and develop iterative optimization algorithms for image reconstruction. Although there has been a large amount of research on model-based iterative reconstruction algorithms, these methods have high time complexity, making real-time image reconstruction difficult. The other approach is purely data-driven deep learning reconstruction algorithms. Thanks to the development of hardware and big data, deep learning methods have also been applied in the field of compressed sensing image reconstruction. Purely data-driven deep learning algorithms utilize the excellent representation capabilities of deep networks to directly learn the mapping between observations and the original image in a data-driven manner, improving reconstruction quality. At the same time, deep networks transfer the reconstruction complexity to the offline training stage, enabling real-time reconstruction with the support of computing hardware. However, these algorithms neglect the mathematical models behind the iterative algorithms, resulting in poor interpretability. Summary of the Invention

[0004] To overcome the shortcomings of slow reconstruction speed and poor interpretability in the two current types of compressed sensing reconstruction algorithms, this invention, based on the idea of ​​deep unfolding, maps the iterative denoising model to a deep learning neural network and uses residual dense blocks as denoising operators, providing an image compressed sensing reconstruction method based on a super gradient denoising network.

[0005] The technical solution of the present invention is as follows:

[0006] A compressed sensing reconstruction method for images based on a super gradient denoising network includes the following steps:

[0007] Step 1: For the original grayscale image, select a row orthogonalized random Gaussian matrix. As a sampling matrix, the entire image is sampled block by block.

[0008] Step 2: Use the initialization matrix to perform image initialization and reconstruction on the low-dimensional sampled measurements, which will then serve as the input for the subsequent iterative denoising and reconstruction module.

[0009] Step 3: Using the initial reconstructed image as input, a denoising network is constructed by cascading a series of residual dense block operators to iteratively denoise the initial reconstructed image. At the same time, a super gradient strategy is introduced to accelerate the iterative convergence rate and obtain the final reconstructed image.

[0010] The beneficial effects of this invention are as follows:

[0011] This invention maps the iterative denoising model to a deep network, uses residual dense blocks as denoising operators for iterative image denoising and reconstruction, and introduces a super gradient strategy to accelerate network convergence. It exhibits faster reconstruction speed and better network interpretability in compressed sensing image reconstruction problems. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the structure of the super gradient denoising and reconstruction network described in this invention; Figure 2 This is the denoising and reconstruction result of a natural image by the super gradient denoising and reconstruction network described in this invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific examples. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of the invention.

[0014] The premise of applying this invention to compressed sensing image reconstruction is that the hypergradient denoising network has been pre-trained. Before describing the specific implementation of this invention in compressed sensing image reconstruction, the training process of the network is briefly described: 88,912 images were randomly extracted from the T91 image dataset. Image patches of size [size missing] are used as the training set. The image sampling compression ratio is set to 30%, generating a [size missing] image patch. Row orthogonalization of random Gaussian sampling matrix .Will Vectorization of image patches vector Using the sampling matrix Further obtain the observation vector Finally, training data pairs are generated. Constructing the error function The Adam optimizer is used to optimize the network parameters, and finally a pre-trained hypergradient denoising and reconstruction network is generated.

[0015] like Figure 1 As shown, an image compressed sensing reconstruction method based on a super gradient denoising network includes the following steps:

[0016] Step 1: For the original grayscale image, select a row orthogonalized random Gaussian matrix. As a sampling matrix, the entire image is sampled block by block.

[0017] Enter 1 sheet The grayscale image was cropped into 64 non-overlapping segments. The image patches are then vectorized to obtain the original image patch vectors. Using row orthogonalization of the random Gaussian sampling matrix Block-by-block sampling is performed to generate low-dimensional sampled measurements. .

[0018] Step 2: Use the initialization matrix to perform image initialization and reconstruction on the low-dimensional sampled measurements, which will then serve as the input for the subsequent iterative denoising and reconstruction module.

[0019] Suppose training data pairs , The original image patch has the following measurement vector: , Let be the number of images in the training set. For all training data, the initialization matrix can be obtained by solving the following least squares problem. :

[0020]

[0021] in express Norm, X = [ x 1 , x 2 , ⋅⋅⋅ , x b ] , Y = [ y 1 , y 2 , ⋅⋅⋅ , y b ] For any measured value y, generate an initial reconstructed image. The process can be represented as:

[0022]

[0023] Step 3: Using the initial reconstructed image as input, a denoising network is constructed by cascading a series of residual dense block operators to iteratively denoise the initial reconstructed image. At the same time, a super gradient strategy is introduced to accelerate the iterative convergence rate and obtain the final reconstructed image.

[0024] The denoising process of the residual dense block denoising operator is as follows: Let the input and output of the residual dense block denoising operator be respectively... and 6 convolutional kernels The function of the convolutional layer is feature extraction, and the output of the m-th convolutional layer can be expressed as:

[0025] F m = H m ([ z ( t ) , F 1 , F 2 , ⋅⋅⋅ , F m − 1 ])

[0026] in, This represents the nonlinear operation of the m-th layer, including the Conv and ReLU activation functions. [ z ( t ) , F 1 , F 2 , ⋅⋅⋅ , F m − 1 ] This indicates that the input and the first The layer output feature maps are concatenated through channels. The residual dense block denoising operator finally employs a convolution kernel... The convolutional layers compress and fuse the extracted feature maps, achieving the fusion of dense features. Furthermore, the input and output of the residual dense block denoising operator are connected through a skip connection, further improving the network's representational power. The final output can be represented as:

[0027] x ^ ( t ) = z ( t ) + H LFF {[ z ( t ) , F 1 , F 2 , ⋅⋅⋅ , F 6 ]}

[0028] in, express Convolution operation.

[0029] The iterative denoising operator for the first residual dense block is expressed as: First, initialize the reconstructed image vector. The first gradient update is performed, and the result of the gradient update is: The result of gradient update The matrix is ​​reshaped and then input into the residual dense block denoising operator for denoising. The output of the denoising process is vectorized to obtain the intermediate denoised result. , The gradient is updated again using the intermediate denoising result. The result of the second gradient update is... The result of gradient update The matrix is ​​reshaped and re-inputted, then fed back into the residual dense block denoising operator for further denoising. The denoised output is then vectorized to obtain the result of the first iteration of denoising. , Following the same method, a total of 9 denoising iterations were performed to finally generate the final result of compressed sensing reconstruction of the image. . Figure 2 This paper presents the reconstruction results of natural images using a super gradient denoising reconstruction network for iterative denoising.

[0030] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above embodiments. Any changes, modifications, substitutions, combinations or simplifications made based on the spirit and principle of the technical solution of the present invention shall be considered equivalent substitutions. As long as they meet the inventive purpose of the present invention and do not deviate from the technical principle of the present invention, they shall fall within the protection scope of the present invention.

Claims

1. An image compressed sensing reconstruction method based on a super gradient denoising network, characterized in that, Includes the following steps: Step 1: For the original grayscale image, select row orthogonalized random Gaussian moments. As a sampling matrix, the entire image is sampled block by block; Step 2: Use the initialization matrix to perform image initialization and reconstruction on the low-dimensional sampled measurements, which will then serve as the input for the subsequent iterative denoising and reconstruction module; Step 3: Using the initialized reconstructed image as input, a denoising network is constructed using a series of cascaded residual dense block operators to iteratively denoise the initialized reconstructed image. Simultaneously, a super-gradient strategy is introduced to accelerate the iterative convergence rate, obtaining the final reconstructed image. The process of the super-gradient strategy can be represented as: Suppose training data pairs , The original image patch has the following measurement vector: , The number of images in the training set. To compress the sensing measurement values; the first t The output of the residual dense block denoising operator is expressed as follows: , No. t The residual dense block iterative denoising operator is represented as follows: Unlike the original iterative denoising model: The specific process of accelerating iterative convergence using the hypergradient strategy can be represented as follows: In the supergradient denoising network, the first t In the next iteration of denoising, the input is first calculated. The gradient at that point, for Gradient updates are performed, and denoising is carried out using the residual dense block denoising operator to obtain intermediate results. Then according to gradient at input Perform gradient updates, and then denoise again using the residual dense block denoising operator to obtain the th... t Output of the next iteration .

2. The image compressed sensing reconstruction method based on a super gradient denoising network according to claim 1, characterized in that: In step 2, the image initialization and reconstruction process can be represented as follows: For all training data, the initialization matrix can be obtained by solving the following least squares problem. : in, express Norm, , ; For any compressed sensing measurement value Generate initial reconstructed image The process can be represented as: 。 3. The image compressed sensing reconstruction method based on a super gradient denoising network according to claim 1, characterized in that: In step 3, the specific process of iterative denoising using residual dense blocks can be expressed as follows: The first t The inputs and outputs of the residual dense block denoising operators are respectively and The first residual dense block m indivual The output of the convolutional layer can be expressed as: in, Indicates the first m The nonlinear operations of the layer include convolution operations and activation function operations; the 6 convolution kernels are... The function of convolutional layers is feature extraction; This indicates that the input and the first The layer output feature maps are concatenated through channels; the residual dense block denoising operator finally uses a convolution kernel as... The convolutional layers compress and fuse the extracted feature maps, thus achieving the fusion of dense features. Furthermore, residual learning is introduced to connect the input and output of the residual dense block denoising operator through a skip connection. The final output of the residual dense block iterative reconstruction can be expressed as: in express Convolution operation.

Images