Compressed sensing reconstruction method and system based on deep learning

Abstract

The invention relates to the technical field of image compressed sensing reconstruction, and discloses a compressed sensing reconstruction method and system based on deep learning, which combine compressed sampling and reconstruction, perform sampling by using a convolutional layer, construct a mapping relation from a measurement matrix to a network layer, parameters in a sampling convolution layer are updated while the system is trained, a more appropriate measurement matrix is obtained through training according to features in a data set, the measurement matrix is made to have self-adaptability and a better image reconstruction effect, the training set is preprocessed, the calculation complexity is reduced, and the sampling efficiency is improved. A convolutional generative adversarial network model is combined in signal reconstruction, and a full connection layer in the network is removed, so that the training speed is higher, and the problem of training over-fitting caused by many parameters of the full connection layer can be prevented. In addition, a measurement value optimization step is added in generator training, and the speed and the performance of signal reconstruction are improved.

Description

technical field [0001] The present invention relates to the technical field of image compressive sensing reconstruction, in particular to a deep learning-based compressive sensing reconstruction method and system. Background technique [0002] Image compressed sensing is a major breakthrough in the field of image processing in recent years. For sparse signals or compressible signals, compressed sensing theory directly encodes the important components containing most of the information in the image signal through a small amount of linear projection to realize the sampling of the original image. , compression and reconstruction. In traditional compressed sensing, when the signal is sampled through the sampling matrix, since the sampling matrix is ​​not related to the signal itself, the sampling matrix must meet the RIP principle to ensure that the key information in the signal is not lost in the sampling process with a high probability. Matrices pose enormous difficulties. A...

AI Technical Summary

Problems solved by technology

This patent uses the product of the measured value and the pseudo-inverse of the measurement matrix as the input signal of the generator, that is, sampling through the measurement matrix. The measurement matrix is ​​preset and cannot adaptively consider the characteristics of the original signal (image). When the data set is trained, it is difficult to ensure a stable sampling effect

Images