Research On Image Recognition For Long-tail Distribution And Cross-domain Data

Convolutional neural networks have achieved excellent performance on image recognition,However,the level of performance needs to be driven by large-scale data,and large-scale and labeled data is very costly and manpower to collect.On the one hand,due to the lack of samples,the dataset in real world approximates long-tailed distribution,which will lead to limited modeling capabilities for tail data.On the other hand,when the number of trainable samples within domain is limited and the lack of labeled samples,the recognition performance will be greatly reduced.In order to solve the above problems,this paper conducts in-depth research from two directions:long-tailed face recognition and cross-domain data image recognition:(1)In the field of long-tail face recognition,low shot learning and long-tail distribution are very difficult problem of data bias.We can solve this problem by learning divergent features and using robust loss functions.We consider enlarging the inter-class variance by directly penalizing weight vectors of last fully connected layer,which represent the center of classes.To the end,we propose Orthogonality loss as an elegant penalty item appends to common classification loss to learn the discriminative representations.More specifically,the optimization objective of Orthogonality loss is moment of cosine similarity of weight vectors.We performed the empirical studies through simulating the long-tail datasets to show the generalization ability of the proposed approach on long-tail distribution datasets.(2)For cross-domain data,Existing methods for domain adaptation focused on the whole domain alignment and class alignment,while ignoring the holistic structure of data and the intra-class compactness of the embedding space.To solve this problem,we present a novel adaptation framework,called Markov Clustering based Graph Convolutional Network(MCGC)for end-to-end training.Based on multi-step Markov random walk on the graph,we devise a novel graph-based loss function via equalizing intraclass transition probability and minimizing inter-class transition probability to supervise the GCN to learn the class discriminative representation.Our method incorporates the graphbased regularization term and GCN to maximize the effective utilization of the graph structure.Experiments on several real-world benchmarks show the proposed approach achieves considerable improvements.