Sample Selection Based Label-Noise Learning Algorithms And Its Application To Facial Attribute Recognition

Noisy labels can significantly deteriorate the performance of learning models in machine learning and computer vision applications.Label-noise learning aims to obtain a robust model from a noisy labeled dataset.Recently,as a kind of intuitive semantic features,facial attributes describe human-understandable visual properties of facial images.The task of facial attribute recognition is to recognize the existence of facial attributes.As a important and hot topic in computer vision,facial attribute recognition has many application scenarios.However,due to the subjectivity and uncertainty of facial attributes,the unknown ratio of noise-labeled samples appears in the facial attribute dataset.The above problem impedes the development of facial attribute recognition seriously.Therefore,how to apply label-noise learning for effective facial attribute recognition is worthy of in-depth study.The main work of this thesis is summarized as follows:First,we propose a novel small-vote sample selection algorithm.Traditional sample selection based label-noise learning algorithms only rely on the information of current batch samples to select clean samples.As a result,these algorithms cannot accurately identify the noise-labeled samples.Therefore,we propose a novel small-vote sample selection algorithm for noisy labels.Specifically,we propose a Hierarchical Voting Scheme,which effectively combines the historical information and current information of the sample as selection criteria.Based on HVS,we further develop Adaptive Clean data rate Estimation Strategy to adaptively estimate the clean data rate by leveraging a 1D Gaussian Mixture Model.Experimental results show that our proposed algorithm has achieved excellent results on several public datasets,verifying the effectiveness of our algorithm.Second,we propose a small gradient norm sample selection algorithm for facial attribute recognition.In order to address the problem of different proportions of noisylabeled samples in the facial attribute dataset,we propose a novel small gradient norm sample selection algorithm for facial attribute recognition.Specifically,we propose a potential noisy-labeled sample selection strategy,which obtains the potential noisy-labeled sample set according to the gradient norm of a sample.Then,based on the batch gradient norm distribution of attributes,the noisy-labeled sample ratios of each attribute are adaptively set,so that different number of noisy-labeled samples of different attributes can be selected,and thus these samples are removed from the training set.A large number of experiments on two public facial attribute datasets have verified the effectiveness of our algorithm.