Quality Assessment of 3D Synthesized Images via Measuring Local Feature Similarity and Global Sharpness

Abstract

Depth-image-based rendering (DIBR) techniques can be used to generate virtual views for free-viewpoint video application. However, the DIBR algorithms will introduce geometric distortions that mainly distribute at the disoccluded regions in the synthesized views. It has been demonstrated that conventional 2-D quality metrics are not suitable for the synthesized views. In this paper, we propose a new quality model for 3-D synthesized images by measuring the block-wise texture similarity and color contrast similarity in critical areas, and the global gradient magnitude deviation. A critical area detection module is first employed using a warping method with morphological operation. Then, the critical areas are partitioned into blocks, which are classified as edge blocks, texture blocks, and smooth blocks by computing discrete cosine transform coefficient values. Block-wise texture similarity and color contrast similarity in the corresponding areas are calculated, which are weighted by the size of critical areas. Furthermore, gradient magnitude deviation is measured to quantify global sharpness. Finally, the two scores are pooled to obtain the overall quality. The experimental results on the IRCCyN/IVC, IETR, and MCL-3-D DIBR image databases indicate that our method achieves higher quality prediction accuracy than the state-of-the-art quality metrics.