Video Quality Assessment Based on Content-Partitioned Multi-Scale Structural Similarity

With the rapid development of video technology, VQA (Video Quality Assessment) has been playing more and more important roles in a variety of modern video processing applications. In this work, we follow a new philosophy in designing VQA algorithm, which uses a content-partitioned image model and motion-based frame weights. First, considering that different regions in image have different perceptual significances, we categorize an image into four different local regions according to the computed gradient magnitude, and allocate different weights to their MSSIM (Multi-scale Structural SIMilarity) scores when pooling. Second, different weights are allocated to different frames according to their different motion conditions. Finally, the overall quality of the entire video sequence is given by combining all the weighted frame values. The algorithm is tested on the LIVE (Laboratory for Image and Video Engineering) video quality database and the results show that it outperforms several modern popular algorithms.