Underwater Target Recognition with Sonar Images Based on Multiple Classifier Systems

Underwater target recognition(UTR) is an important task, where Multiple Classifier System (MCS) has been widely applied. By combining information from multiple sources or features, the improvement of target recognition accuracy can be expected. However, more classifiers in MCS do not necessarily bring better fusion results. In this paper, we focus on two key strategies to improve MCS performance: selecting diverse member classifiers and employing meta-learning methods for fusion. We develop multiple classifiers based on deep features derived from both hand-crafted features of sonar images and original sonar images. Beneficial diversity (BD) is employed to ensure diversity in classifier selection. To implement the fusion of member classifiers, meta-learning methods are used. Experimental results demonstrate the effectiveness of combining diverse classifiers and the advantages of proposed meta-learning methods.