Multi-scale dense description for blind image quality assessment

Bo Liu

Abstract: Blind image quality assessment (BIQA) aims to predict perceptual quality without access to reference images. However, discerning both image content and complex distortions within a unified model poses key challenges. To address this, we propose a single-stage BIQA framework termed multi-scale dense description (MSDD). The key innovation is a dense quality description (DQD) module that matches tailored distortion portrayals to each spatial location using a learned aggregation. This forms fine-grained alignments between contents and distortions. The content and distortion fusion (CDF) module then fuses them via a improved squash function. Finally, the multi-scale quality prediction (MSQP) module recursively combines evidence across stages through cross-scale integration. Extensive experiments show that MSDD achieves state-of-the-art performance and robustness on diverse distortions and contents.