Reproduce “Optimized Learned Image Compression for Facial Expression Recognition” on OpenResearch

About this paper

Xiumei Li, Marc Windsheimer, Misha Sadeghi, Björn Eskofier, André Kaup

Researchers at Friedrich-Alexander-Universität developed an end-to-end framework that optimizes learned image compression for Facial Expression Recognition (FER) using a joint loss function. This approach achieved an 89.12% reduction in compression rate with a 4.04% increase in FER accuracy compared to a baseline, sometimes exceeding uncompressed performance.

Key insights

• Jointly optimizing both the compression and FER models significantly enhances performance, achieving superior compression efficiency and higher recognition accuracy than optimizing compression alone.
• Task-aware compression can lead to a 'feature enhancement' effect, where compressed data sometimes yields slightly higher classification accuracy than uncompressed data, potentially by filtering noise or emphasizing critical features.
• Careful tuning of loss function weights is essential, with observations suggesting that smaller cross-entropy loss weights can sometimes result in better classification accuracy, highlighting complex trade-offs in multi-objective optimization.

Summary by alphaXiv.