Abstract: In recent years, the research on image classification algorithms with noise labels has attracted wide attention in the academic circle. Overfitting To Underfitting(O2U) network is a learning framework for removing noise labels based on the different loss values of noise label samples in Overfitting and Underfitting states. However, this method faces the risk of incomplete clearance of noise label samples. An improved image classification method with noisy labels based on O2U-Net was proposed. By modifying part of the loss function of the denoising frame, the network was robust in the denoised data set, and the influence of O2U-Net on removing unclean noise label samples was reduced. Experimental results show that compared with O2U-Net, the proposed robust loss function combined with denoising framework can improve the classification effect on MNIST, CIFAR-10 and CIFAR-100 data sets. The effects of tag noise rate and noise distribution on classification are summarized. The experiment shows that the classification effect is determined by the noise rate and noise distribution.