Research On Non-linear Metric Learning Algorithms

With the rapid development of network technology, the application of the Internet is more popular. People get a lot of information on the Internet every day. Finding the required information from the Internet is important. The current online information can be broadly divided into two categories, one is based on text information, and the other is the image, audio and other multimedia information. Many scholars have researched on these two kinds of information. At present, it is relatively mature to obtain the required information from a large amount of text. But how to get the information from the image is still an open problem. Metric learning is used to change the original similarity between samples by using the metric function in the sample space. It has important application value in face recognition, identification and so on. Metric learning can be divided into research directions of linear and nonlinear. The basic principle of the linear method is to find the proper distance in all possible Mahalanobis distance by minimizing the objective loss function. The nonlinear algorithm can be divided into several methods based on kernel function, transformation and manifold theory. This paper mainly studies the algorithm of nonlinear metric learning and its application. In this paper, several classical metric learning algorithms are introduced and analyzed. The main works of this paper are as follows:(1) This paper summarizes a number of nonlinear metric learning algorithms based on the transform, and proposes a general framework for this kind of algorithm. On this framework, the concrete perceptron based metric learning and SigmoidML are put forward. These two algorithms are compared with several algorithms which are in the metric learning framework above.(2)This paper constructs the metric tensor and the geodesic distance in the feature space. By optimizing the metric tensor, the approximate optimal geodesic distance is obtained. In order to simplify and speed up the calculation, this work also study the corresponding simplified scheme and fast algorithm, in order to make the algorithm more practical. The comparison experiment shows that the algorithm can effectively improve the experimental results compared with many popular metric learning algorithms.