Learning content-specific codebooks for blind quality assessment of screen content images

摘要

This paper presents a novel blind quality assessment method for screen content images (SCIs) by learning content-specific codebooks. Instead of manually extracting quality-aware features for quality evaluation, the proposed method automatically generates effective features based on a simple feature encoding technique over content-specific codebooks. Considering the mixed content type in SCIs, content-specific codebooks including textual codebook and pictorial codebook are first learned in an offline manner. Given an input SCI, a textual/pictorial segmentation method is first applied to divide the SCI into textual and pictorial regions. Then, patches in different regions are respectively encoded using the learned textual and pictorial codebooks to produce the corresponding feature codes. Finally, the feature codes of each patch are aggregated, by using a percentage-based local pooling scheme, to yield the global feature codes of different regions. The final quality-predictive features used for quality regression are the combined global feature codes of different regions. Despite its simplicity, our method delivers low computational complexity, making it well suitable for real-time applications. Extensive experiments are conducted on three public SCI databases to validate the performance of our method, the results well confirm its superiority over the existing relevant full reference and no reference SCI quality assessment methods. (C) 2019 Elsevier B.V. All rights reserved.