| The deep learning technology driven by massive labeled data has promoted the performance breakthrough and large-scale application of face recognition.Carefully collected and annotated training data can produce large gains in recognition accuracy,model convergence speed,and recognition fairness.However,the current researches in the field of face recognition is mostly limited to topics such as network structure and loss function,and there is little systematical researches on the construction and cleaning of training data.In order to reduce the huge cost of manual annotation,this thesis studies the use of label noise cleaning and automatic clustering algorithms to optimize the quality of the face recognition datasets,reduce the cost of manual annotation,learn to obtain a higher-precision deep recognition model,or optimize downstream tasks such as face generation and editing.At the same time,since the current accuracy of automatic face recognition has surpassed that of humans,it is possible to study the automatic cleaning and labeling of data by machines to obtain more accurate training data than manual labeling.Facing the optimization problem of face datasets,this thesis studies two tasks of image label cleaning and image clustering.The task of label cleaning is to automatically clean out outlier images with label errors from a set of face data,while the task of face clustering is to automatically classify a large set of mixed face images.The main work of this thesis is as follows.(1)In terms of label cleaning,this thesis proposes the FaceGraph label cleaning algorithm based on graph convolutional networks.The network uses the features of face images as vertices,models label cleaning as a graph vertex binary classification problem,trains the graph network with a small number of labeled samples,and uses a cascading two-order graph neural network to perform progressive judgment from global to local scales,obtain accurate vertex classification results,and achieve accurate removal of noise labels.Further,this thesis proposes a meta-learning-based transfer learning cleaning framework AMC,Adaptive Meta Cleaner,which introduces the idea of meta-learning to enable the cleaning model to learn high-quality cleaning knowledge in the source domain,and gradually transfer the knowledge to the target domain.Subsequent experiments show that the proposed method greatly improves the performance of face recognition by data cleaning.The face recognition accuracy of FaceGraph and AMC cleaned WebFace datasets on IJB-C has increased from 68.04%to 72.74%and 73.78%,respectively.(2)In terms of face clustering,this thesis proposes Ada-NETS,Adaptive Neighbour discovEry in sTructure Space,which is an image clustering algorithm that proposes the quality criterion of the input graph of graph neural networks.An adaptive graph building algorithm is also designed to assist the clustering.In addition,this thesis combines the idea of transfer learning to further propose a domain adaptation scheme to extend the performance of the proposed algorithm.On this basis,this thesis proposes UniGraph,a unified graph building clustering framework,which integrates the data distribution of the training set and the target set in the graph building step,and uses the double sampling strategy to improve the performance of the model on the target set.Clustering experiments on MS-Celeb-1M dataset show that a series of data optimization algorithms proposed in this thesis successfully extracts a clean face recognition dataset from noisy environment data,and increases the clustering performance to 93.00%from 91.97%of the state-of-the-art STAR-FC. |
