CTLC: A Co-Training-Based Label Correction Method for Webly-Supervised Learning

Fine-grained image classification aims to recognize fine-grained objects, e.g., different bird species or plane types. However, most proposed methods rely on large-scale annotated data, which is time-consuming and labor-intensive. In recent years, a branch of works propose to learn subtle feature embeddings from webly-annotated images, known as webly-supervised fine-grained recognition. Unfortunately, labels from web are often filled with a large number of noise, which severely hampers the learning of robust neural networks. The most common method at present is to discard high-loss noise samples and select small-loss instances for supervised training, but this selection method still contains a fraction of mislabelled hard samples, which seriously affects the final classification accuracy. In this paper, we propose a novel Co-Training-based Label Correction method (CTLC) to identify noisy samples for robust network learning. Specifically, we train two networks simultaneously to divide images into clean data, in-distribution noisy data, and out-of-distribution data. We regard one’s prediction as the other’s pseudo label for the next step’s training. The out-of-distribution data are discarded and the in-distribution data are relabeled for training. To assure the reliability of re-labeling, a history-assistant in-distribution noise relabeling method is proposed, where cumulative prediction is used to update the labels of the in-distribution noise samples. Extensive experiments on three commonly used noisy fine-grained datasets (i.e., CUB200-2011, FGVC-Aircraft, and Cars-196) show that our method achieves state-of-the-art performance.