| With the rapid development of Internet technology,images show an explosive growth trend in terms of types and quantities.Large scale images have propelled the advancement of deep learning techniques in the field of computer vision.However,accurate datasets provided by experts are expensive and difficult to obtain in the real-world applications.The method of labeling with automated tools and "crowdsourcing" labeling have gradually become the mainstream of obtaining large scale image labels.These low-cost annotation methods will inevitably generate wrong labels,namely label noise.In the field of image classification,label noise will have a negative impact on the classifier,such as reducing the prediction performance of the classifier,increasing the complexity of the model,etc.The impact of the label noise problem is even more severe in large scale datasets,where studies have shown that the tendency of deep neural networks to overfit data makes them easily fit label noise.In response to this problem,this thesis aims to study the classification task of images with label noise.The main research work and research results are summarized as follows:1.A label noise filtering algorithm based on self-supervised learning is proposed.In the existing image classification task,it is a simple and effective way to use the sample selection algorithm to process the label noise.The main purpose of the sample selection algorithm is to select small loss samples from the data as clean samples to guide the model to learn the classification task.However,this method treats hard samples as noise samples,which can improve the classification performance and robustness of the model.In view of the problems existing in the current sample selection algorithm,this thesis uses the mean teacher model and self-supervised learning technology to mine potential hard samples in the data samples discriminated by the model as noise,generate pseudo labels of hard samples,and use them to guide the model to learn image classification task.Experiments show that the label noise filtering algorithm based on self-supervised learning proposed in this thesis can effectively resist the influence of label noise data,and mining hard samples in potential noise can further improve the classification performance of the model,and the anti-noise robustness of the model has reached effects of other advanced label noise filtering algorithms.2.A label noise filtering algorithm based on deep metric learning is proposed.The purpose of the deep metric learning model is to reduce the distance of similar samples in the feature space and increase the distance of different samples.This thesis proposes a label noise filtering algorithm based on deep metric learning.Noisy labels are abnormal points within the same category of data,and the noise data can be discriminated by the similarity between similar data features.On the basis of the deep metric learning algorithm based on memory bank,a label noise sample selection method based on sample similarity is proposed.This method determines whether the sample is clean by calculating the similarity between the sample and the sample data of the same category.And mine hard samples that are not easily discriminated by the model from potentially label noise data.It guides the training of deep metric learning models by constructing positive pairs and negative pairs of clean samples and hard samples,and introducing a supervised contrastive loss.Experiments show that the label noise filtering algorithm based on deep metric learning proposed in this thesis can effectively select clean samples,and the method of constructing positive pairs and negative pairs of difficult samples can further improve the classification ability of the model.The robustness of the model to noise is better than current deep metric learning based label noise filtering algorithms. |
