Research On Label Denoising Method Of Network Image Dataset For Fine-grained Image Classificatio

Fine-grained image classification aims to distinguish subcategories of class,such as the type of bird,and the style of the car.With the development of deep learning,deep convolutional neural networks and their powerful feature representation and generalization capabilities have led to rapid growth in the field of fine-grained image classification.For fine-grained image classification tasks,the labeling of fine-grained datasets usually requires expert knowledge,which makes the task extremely difficult to acquire datasets.Therefore,the direct use of web images for fine-grained image classification has become a popular research direction.However,since the web images have a lot of label noise,training directly with unprocessed web images will greatly degrade the classification performance.To better use web images for fine-grained image classification,this thesis investigates the label denoising method for fine-grained image classification of web images.The main innovations of this article are as follows:(1)Aiming at the problem that the web images collected from the Internet contain label noise,this thesis proposes a new method called Sample Selection based on Fixed Class Center(SSFCC).The method learns the web images by using two identical deep neural networks separately,distributes the image category centers uniformly to the hypersphere and fixes them to better zone the inter-category samples,and selects clean samples for the training of the deep neural network.Finally,experiments on three web-image-based fine-grained datasets reveal 1.9%,0.04%,1.3% performance improvement compared to the state-of-the-art method(i.e.,WSNFG),demonstrating the effectiveness of this approach.(2)Aiming at the problem that the noise rate of real web images is unknown,this thesis proposes a new method called Dynamic Sample Selection based on RANSAC(DSSBR).The method generates the inner point set and dynamically determines the clean sample set for each class by calculating the cosine distance between each sample within the class,and backpropagates the clean sample set.In addition,a cross-entropy loss function based on the central loss constraint is proposed to update the network parameters.The superior classification performance on three web-image-based fine-grained datasets and the high overlap rate of sample selection verify the effectiveness and robustness of this approach.(3)Aiming at the problem that some reusable samples can not be used in web images,this thesis proposes a new method called Combine Sample Select with Loss Correction(CSSLC).Based on the sample selection algorithm(WSNFG),the approach further divides the noise samples in the web images into reusable sample set and noise sample set through the uncertainty of the samples.For the samples in the reusable sample set,this approach takes the label with the most predictions as the true label.Then,this approach combines a reusable sample set and a noise sample set to update the network.Experiments on three web-image-based fine-grained datasets show that performance is boosted by 2.64%,0.9%,4.95%,respectively,compared with the state-of-the-art method(i.e.,SELFIE).