Deep Feature Screening Method by ICT Cascaded with IPSO for Image Recognition

Abstract

Reducing the dimensionality of datasets is considered an important topic addressed in classification problems. In order to reduce the dimension of the features, a new cascaded method is proposed. Firstly, an improved clustering thought (ICT) is used to screen features initially. Secondly an improved particle swarm optimization (IPSO) in which mutation is adopted into the PSO iteration rule is used to filter out the subsets of features whose value of fitness is larger than the certain threshold. Then the support of each feature can be calculated by these selected sunsets. At last, the best feature subset can be screened according to the sorted support. In order to verify the feasibility of this method, 1588 tobacco leaf images belonging to 41 grades have been experimented. And the experiment results show that the proposed deep feature screening method can effectively improve the image recognition rate and recognition speed.