ICASSP 2025 · CCF B · Core B
PEPL: Precision-Enhanced Pseudo-Labeling for Fine-Grained Image Classification
Official project page
TL;DR
PEPL improves semi-supervised fine-grained image classification by generating and refining pseudo-labels with class activation maps, improving label quality for difficult visual categories.
Authors: Bowen Tian†, Songning Lai†, Lujundong Li, Zhihao Shuai, Runwei Guan, Tian Wu, Yutao Yue
Overview
Fine-grained image classification needs reliable supervision over subtle visual details. PEPL improves pseudo-label quality by using Class Activation Maps to generate and refine labels, then applies semantic mixing to strengthen semi-supervised learning.
Key Ideas
- Precision-enhanced pseudo-labels: refine labels with discriminative CAM regions.
- Fine-grained focus: preserve subtle category cues that standard pseudo-labeling can miss.
- Semi-supervised setting: improve performance when only part of the data is labeled.