Compressed sensing video reconstruction method based on dictionary learning residual-error reconstruction

Abstract

The invention relates to the field of compressed sensing and video coding and decoding, and provides a compressed sensing video reconstruction method based on dictionary learning residual-error reconstruction. According to the compressed sensing video reconstruction method, firstly, dividing into groups of pictures (GOP) is performed on a video, wherein a key frame and non-key frames are assigned for each group. The key frame and the non-key frame are successively coded in a frame-by-frame and block-by-block manner according to different sampling rates and different methods. At the decoding end, one GOP is taken, and a multiple-reference-frame weighting averaging is performed for obtaining initial reconstruction; and then an iterative method is adopted for obtaining final reconstruction of the GOP. In iteration, firstly multiple frames are used for performing motion estimation on the current frame for obtaining a motion compensation image; then multiple-reference-frame residual-error domain dictionary learning is performed for obtaining the residual-error domain adaptive group of each block of the current frame, and residual-error reconstruction is performed; and finally, according to the motion compensation image and the residual error of each frame, obtaining the final reconstruction of the GOP, and furthermore obtaining a reconstruction video, thereby realizing high-quality reconstruction of the compressed sensing video. The compressed sensing video reconstruction method can be widely applied on a plurality of fields such as video reconstruction based on compressed sensing.

Description

technical field [0001] The invention relates to the fields of compressed sensing and video reconstruction, and is a compressed sensing video reconstruction method based on dictionary learning residual reconstruction. Background technique [0002] Compressed sensing (Compressed Sensing, CS) is a novel signal sampling mode proposed in recent years, and it has attracted extensive attention from scholars at home and abroad once it was proposed. The traditional Nyquist sampling theorem requires that a signal be sampled at least twice as high as the highest frequency of the signal in order to fully reconstruct the signal. CS breaks this theory by showing that as long as the signal is sparse in some orthogonal space, it is possible to sample a signal with a lower frequency while reconstructing that signal with high probability. The main idea of ​​CS is to simultaneously compress and sample the sparse signal. On the premise of ensuring the information needed to reconstruct the orig...

AI Technical Summary

Problems solved by technology

Although these compressed sensing reconstruction algorithms have achieved certain reconstruction effects and laid a good foundation for the popularization and application of compressed sensing theory, these sparsity prior knowledge is mainly due to the sparsity under some fixed bases, and the adaptability is poor. When solving the decoding and reconstruction of compressed sensing images with variable content, the quality of the reconstructed images is extremely unstable. At the same time, they have not fully exploited and utilized the correlation between image blocks, etc., so the quality of the once-reconstructed image obtained poor

When targeting video files, due to the lack of full use of the temporal correlation between frames, the reconstructed video has low precision and poor quality

Images