296
Asian Journal of Engineering, Sciences and Technology - Volume 12, Issue 1 2022
By Qalb Hussain
Keywords: Computer vision, image processing, compressive sensing and view synthesis.
In this paper, a novel multi-view image compression model is proposed, which leverages block-based Compressive Sensing (BCS) as the encoder and TV-AL3 with view synthesis for decoding. The proposed model encodes a total of N images, out of which two are reference views denoted as VR, and the rest are non-reference views denoted as VNR. The encoding of the reference and non-reference views is performed independently using BCS. However, the sub-rate used for encoding the nonreference views is always lower than the sub-rate used for encoding the reference views. The objective of the proposed model is to improve the reconstruction of non-reference views (VNR) using the reference views (VR). The encoded images produce two observed measurements, which are then independently decoded with TV-AL3 to form two sets of reconstructed views, denoted as V 0 and V 0N R, respectively. After decoding, view synthesis is applied to the reconstructed reference view V 0R to generate a new view denoted as Pl. This new view is then fused with the non-reference views (VNR) using wavelets to produce the projected image B00P. The difference between the measurements of Pl and V NR is then calculated and added to Pl to produce the final reconstructed non-reference view, denoted as B00 NR. Simulation results demonstrate that the proposed model provides superior performance in terms of reconstruction quality for nonreference views (VNR) compared to other CS-based multi-view image compression models. Specifically, on average, the proposed model improves the quality of VNR by 7dB to 13dB at lower subrates.