Learning with Noisy Labels for Image Classification

Abstract

Deep neural networks have achieved remarkable success in image classification tasks, but their performance is known to be vulnerable to incorrect labels, also known as noisy labels. In this paper, we provide a comprehensive overview of robust learning methods designed to mitigate the impact of noisy labels. We first define the concept of noisy labels and discuss the different patterns of label noise. We then introduce real-world datasets that contain noisy labels and outline a promising approach to semisupervised learning for improving robustness against label noise. Finally, we explore the details of methods that enhance robustness to noisy labels, including regularization, label correction, sample selection, and the combination of sample selection with semisupervised learning.