Reproduce “Learning A Sparse Transformer Network for Effective Image Deraining” 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
Xiang Chen, Hao Li, Mingqiang Li, Jinshan Pan
A Sparse Transformer Network (DRSformer) is introduced for single image deraining, which leverages a learnable top-k sparse attention mechanism to filter out irrelevant feature aggregation and incorporates a mixed-scale feed-forward network to capture multi-scale information. The model consistently achieves state-of-the-art quantitative performance, including up to a 0.4 dB PSNR gain on synthetic datasets and superior visual quality on real-world images, by effectively restoring details and removing rain streaks.
Key insights
- Intelligently pruning irrelevant attention scores through a learnable top-k selection significantly improves deraining performance by reducing noise and refining features.
- Explicitly modeling multi-scale information within the feed-forward network is crucial for effectively handling the diverse scales of rain streaks and restoring fine-grained details.
- Combining global modeling with local, multi-scale feature processing, along with an adaptive expert-based refinement, leads to a more robust and higher-quality image deraining solution.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.