Blind Image Quality Assessment via a Hierarchical Perceptual Modulation Network

Abstract

Blind image quality assessment (BIQA) is crucial for perceptual optimization in image processing. To better align objective quality predictions with human perception, researchers have designed BIQA models inspired by the frequency-sensitive characteristics of the human visual system (HVS). However, existing methods usually treat high-frequency (HF) cues as features for direct quality score regression, rather than exploiting them as perceptual modulation signals. In HVS, HF cues play a dynamic role in perceptual modulation, guiding the processing and integration of information across hierarchical stages. Inspired by this, we propose a hierarchical perceptual modulation network (HPM-Net) for BIQA. In the proposed network, we utilize an octave convolution-based high-frequency module (OCHFM) to extract HF cues and then hierarchically modulate them into the quality estimation module through hierarchical feature fusion (HFF). In addition, a contrastive learning-based distortion module (CLDM) is integrated with the modulated features through an attentional feature fusion mechanism. Extensive experiments on five benchmarks show that HPM-Net achieves leading or competitive performance compared to state-of-the-art methods in the majority of datasets with both synthetic and authentic distortions.