MTQ-Caps: A Multi-task Capsule Network for Blind Image Quality Assessment

Yijie Wei, Bincheng Wang, Fangfang Liang, Bo Liu

Abstract: Blind image quality assessment (BIQA) is a field that aims to predict the quality of images without the use of reference images. This field has garnered considerable attention due to its potential applications in areas such as visual understanding and computational imaging. The two primary challenges in BIQA are the differentiation of distortion types and the prediction of quality scores. However, previous approaches typically address only one of these tasks, neglecting the associations between them. In this paper, we propose a multi-task capsule network for BIQA (MTQ-Caps). The proposed MTQ-Caps includes two types of quality capsules routed from CNN-based features: distortion capsules, which correspond to specific distortion types, and content capsules, which encapsulate semantic information for quality score estimation. Consequently, MTQ-Caps utilizes both synthetic distortions and human subjective judgments in a multi-task learning approach. Experimental results demonstrate that our proposed method outperforms state-of-the-art methods on synthetic databases and achieves competitive performance on authentic databases, even without distortion descriptions.