Dual-Band Maritime Ship Classification Based on Multi-layer Convolutional Features and Bayesian Decision

There are some problems arising from the classification of visible and infrared maritime ship, for example, the small number of image annotated samples and the low classification accuracy of feature concatenation fusion. To solve the problems, this paper proposes a dual-band ship decision-level fusion classification method based on multi-layer features and naive Bayesian model. To avoid the occurrence of over-fitting caused by the small number of annotated samples, the proposed method is adopted. First of all, a convolutional neural network (CNN) which has been pre-trained on ImageNet dataset is used and fine-tuned to extract convolutional features of dual-band images. Then, principal component analysis is conducted to reduce the dimension of convolutional feature while L2 normalization is applied to normalize the features after dimensionality reduction. Meanwhile, multi-layer convolutional feature fusion is conducted through the period. In doing so, not only storage and computing resources is reduced, the information of feature representation is also enriched. Finally, a Bayesian decision model is constructed using support vector machine and naive Bayesian theory, for the subsequent dual-band ship fusion classification. According to the experiments results on the public maritime ship dataset, the classification accuracy of the proposed decision-level fusion method reaches 89.8%, which is higher not only than that of the dual-band feature-level fusion by 1.0%–2.0%, but also than that of the state-of-the-art method by 1.6%.