CN-122023574-A - Image reconstruction method of principal component compression domain suitable for adaptive compressed sampling

CN122023574ACN 122023574 ACN122023574 ACN 122023574ACN-122023574-A

Abstract

Embodiments of the present disclosure disclose a principal component compressed domain image reconstruction method suitable for adaptive compressed sampling. The method comprises the steps of carrying out deconvolution reconstruction on the self-adaptive sampling values and the corresponding measurement matrixes to obtain an initial reconstructed image, converting the initial reconstructed image into initial reconstructed features through convolution operation, carrying out K reconstruction stages based on the initial reconstructed features to generate a complete image feature sequence, and carrying out aggregation on the complete image feature sequence to obtain a reconstructed image. This embodiment may improve the efficiency and stability of the reconstruction process.

Inventors

TIAN ZHIFU
MA RUIPENG
YANG SIWEI
HUANG TENGDA
ZHAO WENZHI
HU TAO
WU DI
LIU KAIYUE
LI TINGLI
ZHANG MING
HE RUNZE
ZHU ZAIXING
HAN YI

Assignees

中国人民解放军网络空间部队信息工程大学

Dates

Publication Date: 20260512
Application Date: 20250910

Claims (10)

1. A method of image reconstruction in a principal component compressed domain suitable for adaptive compressed sampling, comprising: performing deconvolution reconstruction on the self-adaptive sampling value and the corresponding measurement matrix to obtain an initial reconstructed image; converting the initial reconstructed image into initial reconstructed features through convolution operation; K reconstruction stages are carried out based on initial reconstruction features to generate a complete image feature sequence, wherein the initial reconstruction features are input of a1 st reconstruction stage, and the complete image features generated in the previous reconstruction stage are input to a next reconstruction stage as initial reconstruction features from a2 nd reconstruction stage; and aggregating the complete image feature sequence to obtain a reconstructed image.
2. The method of claim 1, wherein deconvoluting the adaptive sample values and the corresponding measurement matrix to obtain an initial reconstructed image comprises: For adaptive sampling values by the following formula And corresponding measurement matrix Performing deconvolution reconstruction to obtain an initial reconstructed image : 。
3. The method of claim 1, wherein said converting the initial reconstructed image into initial reconstructed features by a convolution operation comprises: The initial reconstructed image is obtained by the following formula Conversion to initial reconstruction features by convolution operations : , Wherein, the Representing the convolution operation from the image to the feature.
4. A method according to claim 3, wherein the performing K reconstruction stages based on the initial reconstruction features to generate a complete image feature sequence comprises: For a kth reconstruction stage of the K reconstruction stages, performing the following processing steps to generate the complete image feature sequence K-th full image feature in (a) : Based on the first Initial reconstruction features of a reconstruction stage Generating principal component image reconstruction features ; Based on the first Initial reconstruction features of a reconstruction stage Generating compressed domain image reconstruction features ; Reconstructing features from principal component images And compressed domain image reconstruction features Adding and complementing to obtain the complete image characteristics of the kth reconstruction stage 。
5. The method of claim 4, wherein the first based Initial reconstruction features of a reconstruction stage Generating principal component image reconstruction features Comprising: Based on the first Initial reconstruction features of a reconstruction stage Generating principal component image reconstruction features by the following formula : , Wherein, the A convolution operation from the principal component image to the principal component features is represented, Representing the convolution operation from the principal component features to the principal component image, Representing the principal component feature gradient of the kth reconstruction stage, Representing the operation of the near-end mapping, Representing the principal component image of the kth reconstruction stage.
6. The method of claim 4, wherein the first based Initial reconstruction features of a reconstruction stage Generating compressed domain image reconstruction features Comprising: Based on the first Initial reconstruction features of a reconstruction stage Generating compressed domain image reconstruction features by the following formula : , Wherein, the Representing the compressed feature field complementary features of the kth stage, Representing a convolution block operation from the feature domain to the compressed domain, Representing the convolution block operation of a feature from the compressed domain to the feature domain, The characteristic stitching operation is represented as such, Representing the compressed domain feature gradient of the kth reconstruction stage.
7. The method of claim 6, wherein aggregating the complete image feature sequence results in a reconstructed image, comprising: for the complete image feature sequence, the following formula is used Performing aggregation to obtain a reconstructed image : , Wherein, the Representing the convolution operation from the feature to the image.
8. An image reconstruction apparatus adapted for adaptive compressed sampling in a principal component compressed domain, comprising: The deconvolution reconstruction unit is configured to deconvolute the self-adaptive sampling values and the corresponding measurement matrix to obtain an initial reconstruction image; a convolution operation unit configured to convert the initial reconstructed image into an initial reconstructed feature through a convolution operation; The feature reconstruction unit is configured to perform K reconstruction stages based on initial reconstruction features to generate a complete image feature sequence, wherein the initial reconstruction features are input in a1 st reconstruction stage, and the complete image features generated in the previous reconstruction stage are input into a next reconstruction stage as initial reconstruction features from a 2 nd reconstruction stage; And the feature aggregation unit is configured to aggregate the complete image feature sequence to obtain a reconstructed image.
9. An electronic device, comprising: one or more processors; a storage device having one or more programs stored thereon, When executed by the one or more processors, causes the one or more processors to implement the method of any of claims 1-7.
10. A computer readable medium, characterized in that a computer program is stored thereon, wherein the program, when executed by a processor, implements the method according to any of claims 1-7.

Description

Image reconstruction method of principal component compression domain suitable for adaptive compressed sampling Technical Field The embodiment of the disclosure relates to the technical field of image compressed sensing and the technical field of computers, in particular to an image reconstruction method of a principal component compressed domain suitable for adaptive compressed sampling. Background Compressed sensing (Compressive Sensing, CS) is a measurement of signals using undersampled numbers of sample bases using sparsity of the signals, and then reconstructing the original signals as high quality as possible, which has been widely used in different fields such as synthetic aperture radar imaging, magnetic resonance imaging, image encryption and underwater imaging. A uniform sampling compressed sensing (Uniform Compressive Sensing, UCS) model uses the same sampling rate for each block image to obtain sample values and then reconstructs the image. However, the actual image is generally different in complexity of the regional content, i.e. the complex regional image needs to be allocated more samples to be recovered well. Adaptive compressed sensing (Adaptive Compressive Sensing, ACS) adaptively distributes samples according to block image content, can efficiently sense scene content, and has been widely used for single pixel imaging, video snapshot compressed sensing, medical imaging, and THz (Terahertz Imaging, terahertz) imaging. For image reconstruction, benefiting from the success of machine vision, image compressed sensing has developed a pure deep learning network and a deep expansion network (Deep Unfolding Network, DUN). The pure deep learning method utilizes the inverse problem of image reconstruction of the mapping energy mechanics of the neural network, and develops a pure CNN (Convolutional Neural Network ) network, a network based on an attention mechanism and a hybrid network of CNN and a transducer. These networks take advantage of the mapping and generalization capabilities of the network to enhance image de-artifacting and de-noising capabilities, but suffer from well-known black box problems. The depth expansion method combines the traditional optimization algorithm with the neural network, and simultaneously utilizes the strong mapping capability of the network and the measured physical information to be injected into the network to improve the image reconstruction performance. Researchers have fully explored the advantages of conventional optimization algorithms and have now developed networks that develop various conventional optimization algorithms. More, the specific operation of combining the optimization algorithm in the DUN with the network is considered, so that the large measurement of the dimension and information flux of physical injection is proved, and the image reconstruction performance of the DUN can be further released. But typically a large physical implantation dimension means more related operations such as gradient descent operations. In particular, ACS requires more samples for complex areas than UCS, and an increase in the sampling matrix will result in a greater amount of physical injection computation. Thus, efficiency and stability of image reconstruction are reduced. Disclosure of Invention The disclosure is in part intended to introduce concepts in a simplified form that are further described below in the detailed description. The disclosure is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter. Some embodiments of the present disclosure propose an image reconstruction method suitable for a principal component compressed domain of adaptive compressed sampling to solve the technical problems mentioned in the background section above. In a first aspect, some embodiments of the present disclosure provide an image reconstruction method applicable to a principal component compressed domain of adaptive compressed sampling, where the method includes performing deconvolution reconstruction on an adaptive sampling value and a corresponding measurement matrix to obtain an initial reconstructed image, converting the initial reconstructed image into initial reconstructed features through convolution operation, performing K reconstruction stages based on the initial reconstructed features to generate a complete image feature sequence, where the initial reconstructed features are input in a 1 st reconstruction stage, starting from a 2 nd reconstruction stage, inputting the complete image features generated in a previous reconstruction stage as initial reconstructed features to a next reconstruction stage, and aggregating the complete image feature sequence to obtain a reconstructed image. In a second aspect, some embodiments of the present disclosure provide an image reconstruction apparatus suitable for a principal component compressed domain of adapti