Learning From Large-Scale Web Data For Image Classification

Recently,leveraging abundant web data has become a promising strategy to improve the training of convolutional neural network(CNN).However,there are many incorrect tags and noisy data in web images,and many researchers pay attention to alleviate this problem.Specifically,the noisy data is usually judged based on the prediction of basic model which is trained on the standard dataset.Nevertheless,the basic model may be under-fitting,while it also has a preference on samples with same distribution with the standard data and filters out the other useful ones.To address this problem,this paper proposes to iteratively filter out the noisy data as well as fine-tune the CNN model.The proposed method benefits from the growing capability for correcting labels for web images and learning from new data to generate a more effective model.Our contributions are in three-folds.First,this paper proposes a progressive framework to improve the discriminative ability of CNNs and the effectiveness of web image selection.Second,since web images are usually complex while single tag can hardly be accurate,this paper proposes to assign each web image with multiple labels,which mines more training samples to improve the CNN model.Third,to alleviate the problem of different distribution,this paper presents an unsupervised object detection method to process web images and design two criteria to constrain the number,position and label of objects in image.Based on the image-level processing,web data becomes consistent with the distribution of standard data.In the experiments,this paper collects half million web images covering all categories of three public image classification datasets and conduct experiments accordingly,which demonstrates favorable performance against the state-of-the-art methods.