Reproduce “Lightweight Adaptive Feature De-drifting for Compressed Image Classification” 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
Long Peng, Yang Cao, Yuejin Sun, Yang Wang
Researchers at the University of Science and Technology of China developed a Lightweight Adaptive Feature De-drifting Module (AFD-Module) to improve image classification accuracy on JPEG-compressed images by correcting feature distortions directly in the feature space. This computationally efficient and plug-and-play module consistently improved accuracy across various compression levels, particularly on highly compressed data, while maintaining low computational overhead.
Key insights
- Feature drifting caused by JPEG compression across image blocks is strongly correlated with the statistical properties of Discrete Cosine Transform (DCT) coefficients within those blocks.
- Correcting feature degradation in the shallow layers of pre-trained classification models is more effective, as deeper layers contain more semantic information that could interfere with statistical modeling.
- Structural re-parameterization via RepConv blocks enables the model to achieve rich representational power during training while maintaining high computational efficiency during inference by merging branches into a single kernel.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.