Multi-view And Human-Machine Coordination Relation Alignment Metric Learning

With the rapid development of machine learning and computer science,machine learning algorithms have made great progress in the field of computer vision.As a classical machine learning method,distance metric learning has been widely used in many fields of computer vision,such as face recognition,object classification and material classification.With the rapid development of information acquisition technology,data collection capabilities and expressions are becoming more diverse.The explosive growth of data is driving advances in machine learning algorithms,large-scale accurately labeled data greatly increases the upper bounds of various machine learning algorithms and models,but they also increase the difficulty and cost of data labeling work.Through the analysis and integration of information from different perspectives,we can get better results on most data sets.Most current research focuses more on a single perspective of the data.Therefore,the significance of this topic is to explore the regression metric learning from heterogeneous data and how to reduce the cost of large-scale data annotation by human-machine coordination.Aiming at the problem of multi-view and human-machine coordination metric learning,two metric learning methods are proposed in this paper.The methods mainly includes multi-task sparse regression metric learning for heterogeneous classificaton and human-machine coordination metric learning dealing with unlabeled data.The main contributions and innovative work of this paper are shown as follows:1.Referring to the sparse coordination representation of pattern classification,we use dictionary learning method to jointly learn the relation of multi-view data.In the coding stage,features from each view are represented by relevant dictionaries,and with regularization,we enhance the similarity between different features.Through jointly learning,a general matrix M in Mahalanobis distance is obtained to classify the data.Each view effects the classification results in positive or negative ways,so when integrating multi-view data,we can take the distance results into account and get the most optimized classification results.2.The premise of our method is based on a small labeled data set.We use these small labeled data to train a preliminary metric learning model and a support vector machine(SVM)model.When dealing with unlabeled data,we use SVM to select a small part of the hard samples by distance mechanism,and improve the learning model by manually labeling hard samples.Finally,through several iterations,the unknown data can be labeled with a low human cost.