CYFLOD: Cyclic Filtering and Loss Damping for Alleviating Noisy Labels in Fine-Grained Visual Classification

Abstract

We address the challenge of Learning with Noisy Labels (LNL) in fine-grained data sets, a domain exhibiting significant inter-class overlap. Conventional LNL methods fall short in this context. We propose a simple and effective dual-stage approach that can be integrated into any standard transfer learning framework: i) a cyclical iterative filtering scheme in the learning process and, ii) a cyclical loss damping using a SmoothStep function that can be incorporated into any loss function. The proposed integrated scheme iteratively removes noisy labels, enhances data quality, and boosts model generalization. We evaluate our dual-stage solution across diverse data sets, including Stanford Cars and Aircraft for fine-grained categorization, CIFAR-10 for a generic benchmark, and the real-world noise-afflicted Food-101N data set. We conduct our experiments under various noise models, including both symmetric and asymmetric conditions. Our method demonstrates a marked improvement in performance, showcasing its potential in fine-grained classification tasks with noisy labels.