Blind Image Quality Assessment Using Center-Surround Mechanism

Abstract

Blind image quality assessment (BIQA) metrics play an important role in multimedia applications. Neuroscience research indicates that the human visual system (HVS) exhibits clear center-surround mechanisms for visual content extraction. Inspired by this, a center-surround mechanism based feature extraction technique is proposed to solve BIQA problem. The difference-of-Gaussian (DoG) filter, computed in scale-space, has been shown to be able to mimic the center-surround mechanism. In this paper, only DoG maps are employed to characterize the local structure changes in distorted images. The DoG maps are then modeled by generalized Gaussian distribution (GGD) to obtain statistical features. A regression model is learnt to map the features to the subjective quality score. Despite its simplicity, extensive experimental results have demonstrated competitive quality prediction performance and generalization ability of our method.