| With the development of deep learning and computer vision,image classification has been widely used in production and life.In recent years,AI applications such as garbage sorting,product recognition and facial payment have brought great convenience to people’s lives,which have benefited from the development of huge datasets of tagged images.However,the quality of data set labels directly affects the generalization performance and classification accuracy of the model.Therefore,it is particularly important to design a robust and high precision model based on these mislabeled image data sets.The main research work of this paper is as follows:(1)A method of image classification with noisy label based on noise detection is designed.This method combines the noise detection method with the robust loss function method,and reduces the influence of the noise detection algorithm on the experiment by modifying part of the loss function of the noise detection framework.At the same time,through experiments,it is further proved that the conclusion that the lower the noise rate,the better the classification effect is narrow,and the classification accuracy is affected by the noise rate and noise distribution.(2)A collaborative learning method for image classification with noisy labels is proposed.In this method,two networks are used to learn together,and the sample selection is carried out by taking advantage of the characteristics of memorizing clean data first and then noisy data during the training process of images with noisy labels.At the same time,the selected samples are delivered to the peer-to-peer network,which uses robust loss function or label smoothing regularization to calculate the loss,and finally outputs the network model of the data receiver.(3)The two proposed methods are tested on three representative image classification datasets MNIST,CIFAR-10 and CIFAR-100.The experiment shows that the proposed method combining robust loss function with noise detection is superior to other methods in solving the problem of image classification with noise labels.When the data set is CIFAR-100,the noise rate is 60%,and the noise type is uniform noise,the proposed method improves the classification accuracy by 9.32% compared with the method using robust loss function directly and 3.17% compared with the method using noise detection directly.Compared with various commonly used robust loss function methods,the method based on collaborative learning proposed in this paper,on the one hand,improves the classification accuracy,and on the other hand,reduces the degree of model overfitting.When the data set is CIFAR-10,the noise rate is 45%,and the noise type is non-uniform noise,compared with the most advanced robust loss function,the proposed method improves by 5.41%.At the same time,MNIST and CIFAR-10 are superior to the most advanced co-teaching collaborative learning method. |
