Feature Selection For Texture Image Classification

In this paper, features are extracted from underwater acoustic images, then combined with feature selection algorithm to select these features. And to the recognition performance of the target image to select the comparison of several feature selection algorithms, striving for the appropriate selection of texture features acoustic feature selection algorithm.We first extract features from three Underwater Seabed texture images, a total of 20 features of 4 types, the comparison between different feature selection algorithms will base on those features. Then, we introduced the definition and criteria of feature selection and in accordance with the characteristics of algorithms and its follow-up algorithms, feature selection algorithms are sorted as embedded, filter and wrapper. With evaluation and selection criteria, we chose Relief, branch and bound and neural network based algorithm as feature selection algorithms.Relief algorithm uses statistical methods to select relevant features. It is inspired by sample weights algorithm. Its core idea is:good features should make similar features close to each other rather than different types of features. Relief algorithm can only solve the problems with two types, so we use ReliefF to solve the problems with more types.Branch and bound method is the only optimal algorithm. Optimal algorithm is to find the best combinations of features d from the feature set D, criteria J uses to evaluate each feature subset, subset with the largest value J is the best feature subset. The branch and bound method is not convenient and practical in application as for its complicated, so in the paper, an improved branch and bound method (IBB) is chosen to select features.The structure of wrapper, learning algorithm is judged as a feature subset of the black box. The core idea of the algorithm is:learning algorithm and independent evaluation of filter characteristics and follow-up of the classification algorithm will produce a large deviation, while the learning algorithm is based on the performance of the selected feature subset is the better feature of evaluation criteria. In practice, the new search method instead of original space search method of wrapper is proposed to make it more suitable for the requirements of the paper.