Reproduce “Deep Rival Penalized Competitive Learning for Low-resolution Face Recognition” on OpenResearch

About this paper

Peiying Li, Shikui Tu, Lei Xu

Researchers from Shanghai Jiao Tong University introduced Deep Rival Penalized Competitive Learning (RPCL), an adaptation of the classical RPCL principle to deep neural networks, to improve low-resolution face recognition. The method enhanced discriminative power of features by penalizing both target and rival classes, leading to substantial accuracy gains and clearer feature separation across various real-world and synthetic low-resolution datasets.

Key insights

• Explicitly penalizing the 'rival' (most competitive non-target) class, alongside the target class, creates a 'de-learning' force that pushes features away from ambiguous decision boundaries.
• This adaptation of classical Rival Penalized Competitive Learning (RPCL) effectively resolves feature overlap and ambiguity common in low-resolution (LR) face data.
• The dual mechanism simultaneously enhances intra-class compactness and increases inter-class discrepancy, leading to more robust and discriminative feature representations.

Summary by alphaXiv.