Research On Noisy Label Image Classification Algorithm Based On Group-teaching And Visual Semantic Confidence-Guided Mixup

Deep convolutional neural networks(CNNs)have achieved tremendous success in a variety of applications across many disciplines.However,their superior performance relies on correctly annotated large-scale datasets.It is very expensive and time-consuming to get the annotated large-scale datasets.This problem impedes the further expansion of those big datasets.To overcome this limitation,webly supervised learning,which trains networks on images crawled from the Internet by text queries without any human annotation,has been a promising direction recently.The queries key naturally generate labels for webly crawled images,but these labels are highly unreliable and often include a considerable amount of noises.The past research works have demonstrated that the noisy labels could significantly affect the performance of the deep convolutional neural networks on image classification.To combat the drawback,we propose two methods to train a robust convolutional neural network for image classification with noisy labels.The first method is what we call group-teaching.Specifically,we train a group of CNNs simultaneously,and let them teach each other for further training by selecting possibly clean samples for each network in each mini-batch.Group-teaching takes advantage of the diversity of multiple networks and their different learning capabilities,enhancing the ability of each network to identify data with noisy labels,alleviating the impact of noisy labels on the training process,thereby improving the performance of our proposed method.The empirical results on noisy versions of CIFAR-10 and CIFAR-100 datasets demonstrate that our method is superior to the state-of-the-art methods in the robustness for noisy labels.Particularly,to verify the efficacy of our group-teaching in real-world noisy labels distribution,we have also validated the effectiveness of our method on the real-world noisy Web Vision1000-100 dataset.The second method is that we propose a simple yet effective mixup framework for webly supervised image classification.With the guidance of label confidences,the proposed framework suppressed the noisy labels from two perspectives,confidence-biased mixup regularization and confidence-weighted label correction.Specifically,a visual similarity graph whose nodes are images is constructed for each category,and then for each image,meta data accompanied with images in its neighborhood are aggregated to extract semantic information from redundant texts.The experimental results on two large-scale webly supervised learning datasets,including Web Vision1000 and Food101-N,demonstrate the effectiveness of the proposed method.