No-reference image quality assessment based on multi-scale dynamic modulation and degradation information

Image quality assessment is a fundamental problem in image processing, but the complex and varied distortions present in real-world images often affect the model for accurate quality scoring. To address these issues, this paper presents a novel no-reference image quality assessment method based on multi-scale dynamic modulation and gated fusion (MDM-GFIQA), which jointly captures and fuses degradation and distortion features to predict image quality scores more accurately. Specifically, shallow features are first extracted using a pre-trained feature extractor. To explore more deeply perceptual distortion features, we introduce the multi-scale adaptive feature modulation (MsAFM) block into the perceptual network. The MsAFM processes spatial information at different scales in parallel through multiple channels and combines with a multi-branch convolutional block (MBCB), which enables the network sensitive to local features and global information. The comparative learning auxiliary branch (CLAB) is constructed by supervised contrast learning to acquire rich degraded features for guiding the distorted features extracted by the perceptual network. The outputs of these two streams are then merged by our proposed dynamic fusion enhancement module (DFEM), which focuses on key distortion information before passing the fused features to a regression network that predicts the final quality score. Extensive experiments on seven publicly available databases demonstrate the superior performance of the proposed model over several state-of-the-art methods, i.e., achieving the SRCC values of 0.929 (vs. 0.898 in TID2013) and 0.887 (vs. 0.875 in LIVEC).