Reproduce “MSLIQA: Enhancing Learning Representations for Image Quality Assessment through Multi-Scale Learning” 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
Nasim Jamshidi Avanaki, Abhijay Ghildyal, Nabajeet Barman, Saman Zadtootaghaj
MSLIQA introduces a multi-scale learning strategy combined with Test-Time Augmentation to significantly enhance a lightweight No-Reference Image Quality Assessment model. Leveraging a MobileNetV3 backbone, the approach achieves performance comparable to, and often surpassing, more complex state-of-the-art methods on cross-dataset evaluation, notably increasing SRCC by almost 28% over a baseline.
Key insights
- Significant performance gains, competitive with complex state-of-the-art models, can be achieved in NR-IQA through a simple yet effective augmentation strategy on a lightweight backbone.
- Multi-scale learning during training, by exposing the network to diverse resolutions and crop sizes, effectively mitigates information loss and improves discrimination of various distortions.
- Test-Time Augmentation provides additional robustness and generalization by enabling the model to average predictions from multiple perspectives of the test image.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.