In today's era of science and technology, whether scientific or social life is collected and accumulated massive data. Analyzing these massive data is the goal of computer science. And the technique can also promote much of the medicine, geology, meteorology, other traditional disciplines, information biology, combinatorial chemistry and other new cross-discipline. Labeled examples are often time consuming and expensive to obtain, as they require the efforts of human annotators.Semi-supervised learning with few labeled data can obtain a satisfactory performance by integrating unlabeled samples. Semi-supervised learning has a wide range of applications such as video tracking, web page classification, image analysis, intelligent transportation.In this paper, we will study similarity measure, co-training, machine learning using group theory in-depth, applied to image classification and retrieval and have finished the following work:(1) Analyzing the semi-supervised learning algorithms, compare the sameness and difference between various algorithms.(2) Semi-supervised Learning Based on Prior Information Embedding metric has been proposed. According to the position of the data points to select the metric in graph construction.(3) We will focus on the co-training algorithm based on graph structure, and have developed an image retrieval system based on this algorithm.(4) For solving "the subjectivity of human perception" problem, we propose an algorithm based on invariant feature and diversity. |