International Journal of Computer Vision Computer Vision and Pattern Recognition

LEAD++: Unsupervised Fine-Grained Visual Representation Learning with Multi-context Enhanced Entropy-Based Adaptive Distillation

Dong, Jianfeng, Zhang, Zihao, Liu, Daizong, Chen, Shujie, Sun, Jie, Bao, Cuizhu, Yang, Xun, Wang, Meng
Springer Nature · 2026-11-13 · DOI 10.1007/s11263-026-03123-4

Abstract

Fine-grained visual representation learning aims to learn discriminative representations for fine-grained categories within a coarse visual domain. A fundamental challenge in this field lies in distinguishing visually similar categories with subtle inter-class differences, especially when large-scale fine-grained annotations are unavailable. To address this challenge, Unsupervised Fine-grained Visual Representation Learning (FVRL) is introduced, where existing works generally learn representations from raw data, which often struggle to capture subtle inter-class variations due to the limited fine-grained prior knowledge and are easily distracted by semantics-irrelevant background regions, hindering discriminative learning. To alleviate these limitations, we propose LEAD++, an unsupervised FVRL framework that selectively distills fine-grained prior knowledge from a powerful teacher model built upon pre-trained large models. In the teacher branch, large language models (LLMs) generate fine-grained category-aware descriptions to enrich semantic priors, which are incorporated into a frozen vision-language model for guidance. An entropy-based adaptive distillation mechanism dynamically prioritizes high-confidence teacher knowledge while suppressing low-confidence signals. In the student branch, a PCA-based co-localization preprocessing module suppresses semantics-irrelevant regions by extracting dataset-level shared structure. Based on this, a multi-view contrastive learning strategy captures fine-grained variations across viewpoints, enhancing representation consistency under complex conditions. Experiments on three fine-grained datasets demonstrate the effectiveness and efficiency of LEAD++ for unsupervised FVRL. The code for this work is publicly available at https://github.com/HuiGuanLab/LEAD-pp .

Read at publisher →