MASC-IQA: No-Reference Image Quality Assessment Based on Multi-Channel Attention Mechanism via Supervised Contrastive Learning

NR-IQA aims at assessing the perceptual quality of images based on human subjective perception. However, existing NR-IQA methods are likely to miss important fine-grained texture and structural information of key targets, and thus cannot provide reliable quality scores for UAV-captured images in complex power grid inspection scenes. In practical power grid inspection, images are typically captured at high resolutions (e.g. [Formula: see text]K or above) under varying illumination and weather conditions, while quality assessment is required to be performed with low latency on edge or ground-based computing platforms to support downstream inspection tasks. To address these challenges, we propose MASC-IQA, a no-reference image quality assessment model designed for complex background environments. First, we train a Feature Contrastive Module (FCM) on a large-scale real dataset to learn distortion types and degradation levels through supervised contrastive learning, enabling robust distortion-aware feature representation without relying on subjective image ratings during inference. Second, we introduce a Multi-Channel Attention Module (MCAM) to explicitly model inter-channel dependencies and enhance the interaction between global context and local structural details, which is particularly important for preserving perceptually critical information in cluttered industrial scenes. In addition, we release GridVision, a dedicated dataset for NR-IQA in power grid inspection scenarios, consisting of UAV-captured images with realistic distortions annotated by domain experts. Experimental results on several public IQA benchmarks and GridVision demonstrate that MASC-IQA consistently outperforms existing state-of-the-art methods in terms of both prediction accuracy and robustness.