Method for compressed sensing video reconstruction based on recursive convolution neural network

Abstract

The invention proposes a method for compressed sensing video reconstruction based on a recursive convolution neural network, and the method mainly comprises the contents: a compressed sensing network (CSNet), a CSNet algorithm structure, a convolution neural network (CNN), a long-short term memory (LSTM) network, CSNet training, and compressed sensing video reconstruction. The process comprises the steps: extracting movement features through an RNN; extracting visual features through a CNN; carrying out the fusion of the movement features and the visual features; gathering all the extracted features through the LSTM network, and combing the features with a deduced movement in a hidden state to achieve the reconstruction. The method solves a problem that the video reconstruction quality is difficult to guarantee at a high compression rate through a conventional method, and an end-to-end training and non-iteration model is designed. The method improves the compression rate (CR) of a CR camera, improves the video reconstruction quality, reduces the bandwidth of data transmission, and is enabled to support a video application with a high frame rate.

Description

technical field [0001] The invention relates to the field of video compression and reconstruction, in particular to a method for compressive sensing video reconstruction based on a recursive convolutional neural network. Background technique [0002] Video compression and reconstruction are often used in the research of physical and biological sciences, video surveillance, remote sensing technology, social network and other fields. In the research of physical and biological sciences, high-speed cameras are used to record high-speed event characteristics that cannot be recorded by traditional cameras. It can record high-resolution still images of high-speed events, such as an exploding balloon tracking "negligible motion blur and image distortion artifacts." In video surveillance, the area of ​​interest in the surveillance video can be reconstructed, and the image of a specific person or license plate can be enhanced to improve recognition. However, if a camera with a frame ...

AI Technical Summary

Problems solved by technology

[0004] Aiming at the problem that existing methods are difficult to guarantee the quality of video reconstruction under high compression ratio, the purpose of the present invention is to provide a method for compressive sensing video reconstruction based on recursive convolutional neural network, which surpasses the limitations of traditional methods and improves the performance of CS cameras. Compression ratio (CR), and improve the quality of video reconstruction, while reducing the bandwidth of data transmission, making it possible to support high frame rate video applications

Images