MetaBoost: A Novel Heterogeneous DCNNs Ensemble Network With Two-Stage Filtration for SAR Ship Classification
Hao Zheng, Zhigang Hu, Jianjun Liu, Yuhang Huang, Meiguang Zheng
Abstract
Current synthetic aperture radar (SAR) ship classification research mainly focuses on modifying deep convolutional neural networks (DCNNs) and injecting manual features on DCNNs. Yet, the weak robustness of individual models in high-risk scenarios makes it difficult to gain the trust of SAR experts. In this letter, an automated method of heterogeneous DCNNs model ensemble based on two-stage filtration (MetaBoost) is proposed, effectively achieving robustness and high accuracy recognition on SAR ship classification. The principle of MetaBoost is generating a pool of diverse heterogeneous classifiers, selecting a subset of the most diverse and accurate classifiers, and finally fusing meta-features from the optimal subset. MetaBoost is a self-configuring algorithm that automatically determines the optimal type and number of base classifiers to be combined. Extensive experiments on the OpenSARShip and FUSAR-Ship datasets show that MetaBoost significantly outperforms individual classifiers, traditional ensemble models, and feature injection techniques.