ECAIQA: Efficient Cross-Dimensional Attention Network for No-Reference Image Quality Assessment
Year: 2025
Type: conference-paper
Abstract: Existing no-reference image quality assessment (NR-IQA) methods struggle with distortions from generative adversarial networks. We propose ECAIQA, an Efficient Cross-dimensional Attention Network integrating Vision Transformer (ViT) with two novel modules: Convolutional Triplet Attention Module (CTAM) for channel-spatial interaction and Scale Swin Transformer Block (SSTB) for multi-scale global-local feature fusion. A dual-branch architecture aggregates patch-weighted quality scores, achieving state-of-the-art performance across eight benchmarks (LIVE, TID2013, CSIQ, KADID-10K, PIPAL).
Institution: Shanghai University of Engineering Science
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4411800447