On the Performance of Visual Semantics for Improving Texture-Based Blind Image Quality Assessment

Abstract

Blind image quality assessment (BIQA) methods aim to estimate the quality of a given test image without referring to the corresponding reference (original) image. Most BIQA methods use visual sensitivity models, which take into consideration intrinsic image characteristics (e.g. contrast, luminance, and texture) to identify degradations and estimate quality. For example, texture-based BIQA methods are based on the assumption that visual impairments (degradations) alter the characteristics of the image textures and, therefore, their statistics. Although these methods have been are known to provide an acceptable performance, they do not take into account the semantic information of the image. In this paper, we propose a BIQA method that estimates quality using texture characteristics and semantic information. The texture characteristics are obtained using the Opponent Color Local Binary Pattern (OCL) operator. The semantic information is obtained by estimating the probability distribution of the scene characteristics. A random forest regression algorithm is used to map semantic and texture-based features into a quality score. Results obtained testing the proposed BIQA method on several public databases show the method has a good accuracy on quality prediction.