A Data Training Method for Deep Learning Applied to Compressed Sensing Reconstruction

Abstract

The invention belongs to the fields of compressed sensing reconstruction technology and deep learning, and in particular relates to a data training method for applying deep learning to compressed sensing reconstruction. By generating the residual-feature vector pair, a one-to-one correspondence between the residual and the most relevant atom index is established, so that the deep neural network can learn the signal characteristics in the training data, and then apply the deep learning technology to the compressed sensing reconstruction algorithm. Accurately search for the best atoms and improve the accuracy of signal reconstruction.

Description

technical field [0001] The invention belongs to the fields of compressed sensing reconstruction technology and deep learning, and in particular relates to a data training method for applying deep learning to compressed sensing reconstruction. Background technique [0002] With the development of social informatization, the amount of data collected and processed has increased sharply, and the requirements for sensor sampling rate, storage device and transmission bandwidth are getting higher and higher. The traditional signal processing method is to first sample at a high rate and then compress it before storage or transmission. This method will cause a lot of waste of sampled data. As a result, the compressed sensing theory emerged, which can collect signals at a rate much lower than the Nyquist sampling frequency, reconstruct the original signal with high precision, and complete compression while collecting the signal. This theory is widely used in medical signal processing...

AI Technical Summary

Problems solved by technology

[0005] In order to overcome the deficiencies in the above-mentioned prior art, the object of the present invention is to propose a data training method for applying deep learning to compressed sensing reconstruction, construct a residual-feature vector pair model, and guide the pre-training of real data to deep neural networks. Solve the problem that deep learning techniques cannot be directly applied to compressed sensing reconstruction