No-Reference Image Quality Assessment via Local and Global Multi-Scale Feature Integration

In the processes of image capture, transmission, processing, and storage, distortions are inevitable, degrading image quality and affecting both perception and information retrieval. Therefore, assessing image quality is crucial. Current deep learning-based no-reference image quality assessment (NR-IQA) methods face challenges: they fail to preserve both local sensitivity and global details through down-sampling, struggle with multi-scale feature extraction across various distortion types, and lack semantic content perception. Addressing these, we propose the Local and Global Multi-scale Feature (LGMF) algorithm for NR-IQA. This method introduces a grid-based random down-sampling module preserving both local and global details, a pyramid vision Transformer for multi-scale feature extraction, and a global semantic prior extraction module enhancing semantic content perception. Further, a cross-attention feature fusion module selectively integrates these elements, significantly improving quality assessments. The LGMF model also functions as a quality-aware loss function, guiding image restoration networks to achieve superior objective and subjective restoration results. Tested across six datasets, LGMF consistently outperforms existing NR-IQA methods, demonstrating superior generalization and effectiveness in image restoration tasks.