Ship classification based on high-resolution synthetic aperture radar (SAR) imagery plays an increasingly important role in various maritime affairs, such as marine transportation management, maritime emergency rescue, marine pollution prevention and control, marine security situational awareness, and so on. The technology of deep learning, especially convolution neural network (CNN), has shown excellent performance on ship classification in SAR images. Nevertheless, it still has some limitations in real-world applications that need to be taken seriously by researchers. One is the insufficient number of SAR ship training samples, which limits the learning of satisfactory CNN, and the other is the limited information that SAR images can provide (compared with natural images), which limits the extraction of discriminative features. To alleviate the limitation caused by insufficient training datasets, one of the widely adopted strategies is to pre-train CNNs on a generic dataset with massive labeled samples (such as ImageNet) and fine-tune the pre-trained network on the target dataset (i.e., a SAR dataset) with a small number of training samples. However, recent studies have shown that due to the different imaging mechanisms between SAR and natural images, it is hard to guarantee that the pre-trained CNNs (even if they perform extremely well on ImageNet) can be finely tuned by a SAR dataset. On the other hand, to extract the most discriminative ship representation features from SAR images, the existing methods have carried out fruitful research on network architecture design, attention mechanism embedding, feature fusion, etc. Although these efforts improve the performance of SAR ship classification to some extent, they are usually based on more complex network architecture and higher dimensional features, accompanied by more time-consuming storage expenses. Through the analysis of SAR image characteristics and CNN feature extraction mechanism, this study puts forward three hypotheses: (1) Pre-training CNN on a task-specific dataset may be more effective than that on a generic dataset; (2) a shallow CNN may be more suitable for SAR image feature extraction than a deep one; and (3) the deep features extracted by CNNs can be further refined to improve the feature discrimination ability. To validate these hypotheses, we propose to learn a shallow CNN which is pre-trained on a task-specific dataset, i.e., the optical remote sensing ship dataset (ORS) instead of on the widely adopted ImageNet dataset. For comparison purposes, we designed 28 CNN architectures by changing the arrangement of the CNN components, the size of convolutional filters, and pooling formulations based on VGGNet models. To further reduce redundancy and improve the discrimination ability of the deep features, we propose to refine deep features by active convolutional filter selection based on the coefficient of variation (COV) sorting criteria. Extensive experiments not only prove that the above hypotheses are valid but also prove that the shallow network learned by the proposed pre-training strategy and the feature refining method can achieve considerable ship classification performance in SAR images like the state-of-the-art (SOTA) methods.
