Abstract
For fine-grained visual classification (FGVC), the annotation of fine-grained data usually requires expert knowledge, which makes it difficult to obtain the fine-grained data. Recently, an increasing number of researchers start to pay attention to using web images for fine-grained classification directly. However, noisy labels inevitably exist in web images. Noisy labels are harmful to Convolutional Neural Networks (CNNs) due to their strong capability in fitting any data. To this end, we propose an end-to-end label denoising based fine-grained classification method to tackle noisy labels. We separate the clean samples and the noisy samples based on the cosine similarity between the sample and the class center. For samples with low cosine similarity to the class center, we select the reusable samples with prediction uncertainty and correct the labels of reusable samples to the true labels. Finally, the clean samples and the corrected reusable samples are fed into the network for further training. We provide extensive experimental results to demonstrate the superiority of the proposed approach. The data and source code of this work have been made available at: https://github.com/huarm123/LDWFC.git.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call