Feature Selection For Multi Label Image Data

With the development of science and technology, the ability of people to collect information is more and more strong, so a large number of high dimensional data has been accumulated. For example, Facebook,label-me, flicker, Tencent space and other sites can produce the image data which is common among the high dimensional data. Owing to the need for image data management, it turns into an important research topic about the automatic understanding of image content. The research of this topic mainly involves the prediction of image labels. In particular, it is the optimal selection of the image representation features, in order to improve the accuracy of the existing prediction methods.Feature selection refers to the process to select the best feature subset from the feature space according to a certain standard, our goal is to retain the characteristics of classification of value representative,according to the contribution in classification to select some features constitute a subset of features who can best represent all the characteristics, so as to reduce the dimensionality of the feature space and the data, improve the performance of the pattern recognition algorithm in label prediction and image recognition.At present, according to the characteristics of high dimensional image data selection problem research is still relatively small. The problem is mainly reflected in the dimension of image features is too high,the traditional method has a character of large computation and high storage consumption can not be proceed smoothly. In order to solve this problem, this paper based on the feature similarity method of grouping, to avoid the high dimension feature as a whole, directly for feature selection,but the group feature score in the feature grouping, then each feature score in one group in the organic integration of each feature in all features of the score, to achieve robust feature selection. Prediction of EPS games through the multi-label image standard in the world the data set, corel5k,and iaprtc12 for the full experiment proves the effectiveness of the proposed algorithm.