Integrate Traditional Hand-Crafted Features into Modern CNN-based Models to Further Improve SAR Ship Classification Accuracy
Tianwen Zhang, Xiaoling Zhang
Abstract
Current most state-of-the-art convolutional neural network (CNN) based synthetic aperture radar (SAR) ship classifiers have achieved a great accuracy progress compared with traditional hand-crafted feature-based methods. Nevertheless, these existing CNN-based models almost all uncritically abandon traditional mature hand-crafted features, but heavily rely on abstract features extracted by deep networks, which possibly brings some challenges in further improving classification performance. Therefore, to address this problem, in this paper, we propose a novel approach of integrating traditional hand-crafted features into modern CNN-based models so as to further improve SAR ship classification accuracy. To be specific, we respectively integrate four types of traditional hand-crafted features, i.e., grey features, canny edge features, corner harris features, and histogram of oriented gradient (HOG) features, into four types of classic and representative CNN-based models, i.e., AlexNet, VGGNet, ResNet and DenseNet, successfully. Experimental results on the open OpenSARShip dataset indicate that after traditional feature integration, the SAR ship classification performance of CNN-based models can achieve a huge progress, which also shows that it is not advisable if uncritically abandoning traditional mature hand-crafted features, because, in fact, traditional features can also play an important role in CNN-based models.