Search

CN-115953289-B - Light-weight image reconstruction method based on AMP network

CN115953289BCN 115953289 BCN115953289 BCN 115953289BCN-115953289-B

Abstract

S1, acquiring a two-dimensional image, dividing the two-dimensional image into image sub-blocks with a preset number and a fixed size through a block function, and converting the image sub-blocks into one-dimensional vector signals; S2, constructing a light-weight AMP network, wherein the light-weight AMP network comprises a sampling frame, an initialization module and a multi-layer reconstruction block which are sequentially connected, and S3, inputting one-dimensional vector signals into the light-weight AMP network to output a reconstructed image. According to the invention, the light-weight AMP network is constructed, full-image convolution is introduced into the multi-layer reconstruction block to fully utilize image global information, so that the artifact is removed, an additional artifact removal module is not required to be introduced, the reconstruction time and model parameter quantity are greatly reduced on the premise that the reconstruction quality is quite even better, and the resources and time required by image reconstruction are reduced.

Inventors

  • ZHANG JUN
  • GUO ZHIJING

Assignees

  • 广东工业大学
  • 人工智能与数字经济广东省实验室(广州)

Dates

Publication Date
20260512
Application Date
20221216

Claims (8)

  1. 1. The light-weight image reconstruction method based on the AMP network is characterized by comprising the following steps of: s1, acquiring a two-dimensional image, dividing the two-dimensional image into image sub-blocks with a preset number and fixed size through a block dividing function, and converting the image sub-blocks into one-dimensional vector signals; S2, constructing a light-weight AMP network, wherein the light-weight AMP network comprises a sampling frame, an initialization module and a multi-layer reconstruction block which are sequentially connected; s3, inputting the one-dimensional vector signals into a lightweight AMP network to output the reconstructed image; The multi-layer reconstruction block processing process comprises the following steps: inputting the initialized image signal output by the initialization module into a multi-layer reconstruction block to reconstruct an image in an iterative manner, and the first The first image sub-block The iterative process of the layer is: (3) (4) Wherein, the And Respectively represent the first Measurement value and the first image sub-block The output of the layer is provided with, And Respectively the first Nonlinear iterative functions and step sizes of layers; Will be And Substituted into The following steps are: (5) in the above formula (5), the original signal is In which only a sufficiently good fit is required A sum primary signal can be obtained A sufficiently similar reconstructed signal; Using Fitting to Wherein For iterative values In the form of a two-dimensional image signal, For a four-layer convolution network, the convolution kernel size and the filling size of each layer are respectively as follows And The first three convolutions have offset and ReLU layers, and the last convolutions have no offset and ReLU layers; On the heald, the first The process of layer iteration can be summarized as: (6) (7) Will be the first All image sub-blocks of the layer output are spliced into a complete image As a means of Input of (1), then The layer iteration is expressed as: (8) (9) (10) Wherein the method comprises the steps of Is that Image sub-blocks An integrated matrix; To further reduce the iterative matrix computation, substituting (8) into (10) and deriving as follows: (11) Wherein the method comprises the steps of ; The multi-layer reconstruction block parameters include a sampling matrix Step size of each layer And Parameters of a convolutional network ; The mean square error is adopted as a multi-layer reconstruction block loss function, so that the network can reduce the original image signal better after training And reconstructing an image signal Error between: (12) Wherein, the For the number of pictures of the training set, Is the first Training image signals.
  2. 2. The light-weight image reconstruction method based on the AMP network according to claim 1, wherein the specific process of step S1 is as follows: acquiring a two-dimensional image, dividing the two-dimensional image into L non-overlapping images with the size of Image subblock { of (2) Wherein Converting image sub-blocks into one-dimensional vector signals using vectorization functions Wherein Will be Integrated into a matrix form and marked as 。
  3. 3. The method for reconstructing a light-weight image based on an AMP network according to claim 1, wherein the one-dimensional vector signals are represented by a sampling matrix in a sampling framework Performing linear mapping to complete compressed sampling, wherein the expression of the linear mapping is that The sampling process is represented as follows: (1) Representing a sampling matrix , Representing the sampled value of the sample, 。
  4. 4. The light-weight image reconstruction method based on the AMP network according to claim 1, wherein the initializing module receives the sampled values from the sampling frame and initializes the sampled values to obtain an initial image signal, and the expression is as follows: (2) Wherein, the Sampling value, initial image signal , For sampling matrix Is a transpose of (a).
  5. 5. A light-weight image reconstruction system based on an AMP network is characterized by comprising a memory and a processor, wherein the memory comprises a light-weight image reconstruction method program based on the AMP network, and the light-weight image reconstruction method program based on the AMP network realizes the following steps when being executed by the processor: s1, acquiring a two-dimensional image, dividing the two-dimensional image into image sub-blocks with a preset number and fixed size through a block dividing function, and converting the image sub-blocks into one-dimensional vector signals; S2, constructing a light-weight AMP network, wherein the light-weight AMP network comprises a sampling frame, an initialization module and a multi-layer reconstruction block which are sequentially connected; s3, inputting the one-dimensional vector signals into a lightweight AMP network to output the reconstructed image; The multi-layer reconstruction block processing process comprises the following steps: inputting the initialized image signal output by the initialization module into a multi-layer reconstruction block to reconstruct an image in an iterative manner, and the first The first image sub-block The iterative process of the layer is: (3) (4) Wherein, the And Respectively represent the first Measurement value and the first image sub-block The output of the layer is provided with, And Respectively the first Nonlinear iterative functions and step sizes of layers; Will be And Substituted into The following steps are: (5) in the above formula (5), the original signal is In which only a sufficiently good fit is required A sum primary signal can be obtained A sufficiently similar reconstructed signal; Using Fitting to Wherein For iterative values In the form of a two-dimensional image signal, For a four-layer convolution network, the convolution kernel size and the filling size of each layer are respectively as follows And The first three convolutions have offset and ReLU layers, and the last convolutions have no offset and ReLU layers; On the heald, the first The process of layer iteration can be summarized as: (6) (7) Will be the first All image sub-blocks of the layer output are spliced into a complete image As a means of Input of (1), then The layer iteration is expressed as: (8) (9) (10)(10) Wherein the method comprises the steps of Is that Image sub-blocks An integrated matrix; To further reduce the iterative matrix computation, substituting (8) into (10) and deriving as follows: (11) Wherein the method comprises the steps of ; The multi-layer reconstruction block parameters include a sampling matrix Step size of each layer And Parameters of a convolutional network ; The mean square error is adopted as a multi-layer reconstruction block loss function, so that the network can reduce the original image signal better after training And reconstructing an image signal Error between: (12) Wherein, the For the number of pictures of the training set, Is the first Training image signals.
  6. 6. The light-weight image reconstruction system based on an AMP network according to claim 5, wherein the specific process of step S1 is: acquiring a two-dimensional image, dividing the two-dimensional image into L non-overlapping images with the size of Image sub-blocks of (a) Wherein Converting image sub-blocks into one-dimensional vector signals using vectorization functions Wherein Will be Integrated into a matrix form and marked as 。
  7. 7. The light-weight image reconstruction system based on an AMP network according to claim 5 wherein the one-dimensional vector signals are represented in a sampling frame using a sampling matrix Performing linear mapping to complete compressed sampling, wherein the expression of the linear mapping is that , Representing a sampling matrix , Representing the sampled value of the sample, 。
  8. 8. A computer-readable storage medium, wherein an AMP network-based lightweight image reconstruction method program is included in the computer-readable storage medium, which when executed by a processor, implements the steps of an AMP network-based lightweight image reconstruction method according to any one of claims 1 to 4.

Description

Light-weight image reconstruction method based on AMP network Technical Field The present invention relates to the field of image reconstruction technology, and more particularly, to a light-weight image reconstruction method and system based on an AMP network, and a computer readable storage medium. Background With the rapid development of the internet, high-definition pictures, videos, moving pictures and other large-capacity unstructured data are widely generated, and in addition, people continuously pursue the image quality effect, extremely high requirements are brought to channel bandwidth, processing speed and memory resource occupation in the transmission process, while image compression sensing (Compressive IMAGE SENSING, CIS) is a technology capable of sampling and reconstructing image signals at an ultra-low sampling rate, and due to a simple coding mechanism, CIS is usually used for image information acquisition and data compression in a resource-limited scene, but for a decoding end, with the rapid development of mobile application, more and more resource-limited scenes require as few resources and time as possible for image reconstruction, so that the complexity, model scale and reconstruction time of a reconstruction algorithm are required to be reduced as much as possible on the premise of not sacrificing reconstruction quality. Compressed sensing (Compressive sensing, CS) refers to a signal processing technique that performs compressed sampling and reconstruction of a signal at an ultra-low sampling rate, and cande s, romberg, tao and Donoho et al, 2006, indicate that when a signal has a sparse representation, it can be reconstructed accurately by a series of linear, non-adaptive observations. For the K-stage sparse signal x e R N, in the encoding end we can use a linear mapping y=Φx to compress and sample the signal at the same time, and obtain the observed value y e R N, where Φe R M×N (M < < N) is the sampling matrix. Then we can transmit the extremely short observation value y to the decoding end, and after the decoding end obtains y, x can be accurately reconstructed through the corresponding algorithm. Although natural image signals are not sparse, we can wavelet transform the image signals to obtain an approximation coefficient representation of the signal. The wavelet transform converts the image signal into high and low frequency parts, the low frequency part gives a rough scale approximation of the image, and the high frequency part fills in the details of the image, when we calculate the wavelet coefficients of a natural image, most of the coefficients are very small, so we can set the small coefficient part to zero, and an approximate representation of the image can be obtained. Therefore, for a natural image signal s e R N, the measured value y e R M is obtained by sampling y=Φs through the sampling matrix Φe R M×N, at this time we can reconstruct the approximate sparse representation x e R N of the image by y, and inverse wavelet transformAn image signal is reconstructed, where ψ e R N×N is the inverse wavelet transform matrix. The existing schemes mainly can be divided into two types, namely a Non-iterative neural network (Non-ITERATIVE NEURAL NETWORK), an initial neural network is firstly built to map y to x, then the strong learning capability of deep learning (DEEP LEARNING, DL) is utilized, a high-quality neural network and a better sampling matrix are trained through a large amount of data, the reconstruction speed of the schemes is high, the image reconstruction quality is high, but the schemes are regarded as a 'black box' in the training process, the mathematic interpretation of the schemes is weak, and a sufficient amount of data is needed for training. The second type is a deep non-folding algorithm (deep unfolding methods), the scheme expands the traditional compressed sensing iterative algorithm into a neural network, so that the method has the advantage of strong mathematical interpretation of the traditional iterative algorithm, can inherit excellent learning ability of the neural network, and utilizes training data to obtain a signal reconstruction model with high speed and high reconstruction quality. Because the sampling matrix parameter, the sampling calculation amount and the reconstruction calculation amount are very large when a whole image is subjected to compressed sampling, in order to reduce the calculation amount, the scheme firstly needs to divide the image into a plurality of non-overlapping image sub-blocks, then respectively samples each sub-block, and the reconstruction of each image sub-block is independently carried out in the reconstruction process, so that a large amount of artifacts (Blocking artifacts) are generated, an additional artifact removal module is often required to be introduced after the block-by-block reconstruction, the calculation complexity and the parameter amount of the model are greatly increased, the storage space and th