Research On Robustness Learning Methods Under Noisy Labels

In recent years,with the continuous development of computer vision,image classification applications has been widely used in people’s daily life,such as face recognition,vehicle recognition and so on.However,with the explosive growth of image data on the Internet,a large amount of data is collected directly from crowdsourcing systems,search engines and social networking sites,would possibly cause a lot of noisy label.It is almost impossible to provide accurate manual annotation for all datasets.Deep learning relies on large amounts of trusted data,but the noise prevails in real situations can be detrimental to the performance of the algorithm.Using the image data with noisy labels to learn robustly has therefore become a topic of increasing interest.Although some progress has been made in research related to the problem of robust classification of noisy data,there are still many problems that need to be solved.1)The single-label classification with noisy label tasks:The deterministic model(mostly built as re-weight method or loss correction)usually neglect the sample ambiguity.And some meta-based methods do not make good use of the structural information of the output of the classification network.2)Partial multi-label classification tasks:Some methods require matrix computation based on global data,which is inefficient or even difficult to calculate in the case of large dataset.In terms of the utilization of data information,traditional methods usually would not utilize the structural information of label space and feature space effectively,which may lead the model to overfit the noisy labels in the candidate label set.This paper presents an in-depth study of robustness methods for image classification with noisy labels,and proposes two robust methods to address some problems of existing studies.The main contributions of this paper are as follows:1)For the single-label classification with noisy label tasks:In this paper,we propose a robustness algorithm based on meta-learning with adaptively rectify training.We model the learning process as a hierarchical probabilistic model and consider the rectifying vector as a latent variable,and propose a warped probabilistic inference(WarPI)to achieve an effective estimation of the prediction posterior.It helps to solve the problem of sample ambiguity and further improve the generalization of the proposed model.We have conducted extensive experiments on several datasets for different types of noise.All results demonstrate the superiority of our proposed WarPI while compared to other models in terms of effectiveness,efficiency and stability.2)For the partial multi-label classification tasks:we propose an end-to-end partial multi-label learning framework Co-GCN,which is based on the graph convolution network.the framework construct two graph structures based on the label space and feature space separately,which avoids complex arithmetic operations while further improving the robustness of the model via the full utilization of data structure information and label statistics.The proposed method is compared with some state-of-the-art methods on two multi-label image datasets to evaluate the effectiveness,and the results demonstrate the superiority of the proposed method outperforms all the compared methods.