PTLO: A model-agnostic training strategy based on progressive training and label optimization for fine-grained image classification

Abstract

Compared to conventional image recognition, fine-grained classification exhibits increased vulnerability to labeling noise due to the presence of closely related categories, resulting in degraded performance on complex and non-representative samples. While existing approaches mitigate these issues through data cleaning, loss modification, and semi-supervised learning techniques, they often overlook the intrinsic attributes within training samples. Instead of designing any network architectures, this study introduces a model-agnostic progressive training strategy comprising of progressive training and label optimization, where the former is to decrease the affect from the noisy samples by facilitating a graduated learning approach in an easy-to-hard manner, while the latter is to denoise the label noises. Theoretical analysis also demonstrates that the proposed method uncovers valuable cues hidden in the training data, thereby enhancing the robustness of any learning-based models. Experimental evaluations on fine-grained classification benchmarks (e.g., CUB-200-2011, DTD, and Food-101) across various mainstream classification networks demonstrate the effectiveness of our training strategy. Code is available at https://github.com/cb-rep/LPPT.