Reproduce “MANIQA: Multi-dimension Attention Network for No-Reference Image Quality Assessment” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, Yujiu Yang
MANIQA addresses no-reference image quality assessment for restored images by combining attention across feature channels and image regions. It predicts each patch’s quality and importance, then aggregates patch scores. In the NTIRE 2022 challenge, its ensemble ranked first in the final private test, with SROCC 0.704 and PLCC 0.740; these results measure agreement with ratings, not universal taste.
Key insights
- Leveraging Vision Transformers, particularly through multi-layer feature concatenation, provides a richer understanding of image distortions by capturing diverse abstraction levels.
- Combining channel-wise global attention (Transposed Attention Block) with spatial-wise local attention (Scale Swin Transformer Block) effectively captures comprehensive image information for complex degradations.
- A dual-branch prediction system that assigns both a quality score and an importance weight to image patches better aligns with human visual perception by balancing salient areas and areas of higher quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.