Compressive-sensing recovery of images by context extraction from random samples

Abstract

Image Compressive Sensing (CS) provides a scheme of low-complex image coding, but coping with the recovery quality has been a challenge. Even the excessive investment of computations into recovery cannot prevent the quality degradation due to the lack of appropriate allocation for sampling resources. In light of this, this paper fuses a context-based allocation into image CS in order to improve the recovery quality with fewer computations. Independent of original pixels, the context features of blocks are extracted from random CS samples. According to the block-based distribution on context features, more CS samples are allocated to non-sparse regions and fewer to sparse regions. The proposed context-based allocation enables a linear recovery model to accurately recover images. The contributions of this paper include: (1) an adaptive allocation involving the context features extracted from CS samples, (2) a padding Differential Pulse Code Modulation (DPCM) to quantize the adaptive CS samples, and (3) a regrouping module to improve the quality of linear recovery. Experimental results show the proposed image CS system objectively and subjectively improves the recovery quality of an image while guaranteeing a low computational complexity, e.g., it achieves average 30.85 dB PSNR value on the five $512\times 512$ test images, and costs about 10 seconds on a computer with 3.30 GHz CPU and 8 GB RAM. Besides, the proposed system presents a competitive performance to the recent deep-learned image CS systems.