Alleviating Noisy-label Effects in Image Classification via Probability Transition Matrix

  • 2021-10-19 16:56:14
  • Ziqi Zhang, Yuexiang Li, Hongxin Wei, Kai Ma, Tao Xu, Yefeng Zheng
  • 0

Abstract

Deep-learning-based image classification frameworks often suffer from thenoisy label problem caused by the inter-observer variation. Recent studiesemployed learning-to-learn paradigms (e.g., Co-teaching and JoCoR) to filterthe samples with noisy labels from the training set. However, most of them usea simple cross-entropy loss as the criterion for noisy label identification.The hard samples, which are beneficial for classifier learning, are oftenmistakenly treated as noises in such a setting since both the hard samples andones with noisy labels lead to a relatively larger loss value than the easycases. In this paper, we propose a plugin module, namely noise ignoring block(NIB), consisting of a probability transition matrix and an inter-classcorrelation (IC) loss, to separate the hard samples from the mislabeled ones,and further boost the accuracy of image classification network trained withnoisy labels. Concretely, our IC loss is calculated as Kullback-Leiblerdivergence between the network prediction and the accumulative soft labelgenerated by the probability transition matrix. Such that, with the lower valueof IC loss, the hard cases can be easily distinguished from mislabeled cases.Extensive experiments are conducted on natural and medical image datasets(CIFAR-10 and ISIC 2019). The experimental results show that our NIB moduleconsistently improves the performances of the state-of-the-art robust trainingmethods.

 

Quick Read (beta)

loading the full paper ...