Noisy Label Learning Methods Via Sample Selection

In recent years,deep neural networks have achieved impressive success in image classification,semantic segmentation and many other tasks.At present,the success of deep neural networks generally relies on a large number of high-quality labeled samples.In other words,the labeling information of training example is assumed to be accurate and complete.However,it is extremely expensive and time-consuming to label high-quality datasets,thus deep models are usually trained on data with lots of corrupted labels.Noisy label learning via sample selection aims to distinguish clean samples and noise samples from noisy datasets,and then utilizes them separately to improve the robustness of the model.In this paper,we investigate the problem of noisy label learning via sample selection and propose two novel methods accordingly.On one hand,numerous researches have proved that deep neural networks can fit all training data in the end even given data with noisy labels,and thus result in poor generalization performance.In this paper,a novel noisy label learning method via Dynamic Loss Thresholding(DLT)is proposed.During the training process,through recording the loss value of each sample and calculating dynamic loss thresholds,DLT compares the loss value of each sample with the current loss threshold and extracts clean samples for the next step of training.Experiments on several image classification datasets demonstrate substantial improvements over recent state-of-the-art methods.On the other hand,by introducing self-training semi-supervised learning techniques,a novel noisy label learning method named Semi DLT is proposed to accommodate unlabeled data exploitation.The proposed method discards the corrupted labels of potentially noisy samples and regards them as unlabeled data.After that,Semi DLT generates pseudo-labels for the unlabeled data,and utilizes the Mixup techniques for data augmentation.The loss function is instantiated by integrating cross entropy and mean square error.Experimental results on several benchmark datasets show that the proposed method can take advantage of unlabeled data effectively and improve the classification performance.This paper consists of four chapters.The first chapter introduces the background,related work and the problems to be solved.The second chapter introduces the noisy label learning method named DLT via dynamic loss thresholding.The third chapter introduces the noisy label learning method Semi DLT via unlabeled data exploitation.The fourth chapter summarizes the thesis.