Robust Convolutional Neural Network For Label Noise And Imbalanced Data

In recent years,deep learning based convolutional neural network achieves significant success in computer vision,deep learning becomes the most representative and effective technology in computer vision.But deep learning depends on the large-scale well-labelled and class-balanced data.It is extremely difficult to obtain large-scale welllabelled and class-balanced data.For face recognition and image classification,how to make use of the data with label noise or imbalanced data when training convolutional neural networks are challenging problems to be resolved.In this paper,problems of data with label noise and imbalanced data are deeply studied.My works in this paper are as follows:1)Propose self-paced robust face recognition method for large-scale dataset with different types and different ratio label noise,and the method works on dataset suffer from corrupted labels and outliers.2)Propose class-balanced re-sampler and loss method for imbalanced data,this method improves the accuracy of image classification and reduce the training time.According to experiment results,above-mentioned methods makes convolutional neural networks achieve significant performance on large-scale dataset with label noise and imbalanced dataset.