Asymmetric and Square Convolutional Neural Network for SAR Ship Detection from Scratch
Long Han, Xi Bei Zhao, Wei Ye, Da Ran · 2020
Despite there have been little research on synthetic aperture radar (SAR) ship detection from scratch, the transfer learning-based studies are still the mainstream. Due to some limitations of the fixed structures of the backbone, it is hard to improve and optimize these parts of networks and there is between domain mismatch which restricts the performance of the convolutional neural network (CNN) to some extent. Addressing the issue, we design an asymmetric and square convolution block (A-S CB) which is easy to be embedded into any CNNs and helps to significantly reduce the number of parameters and computations without serious damages to detection accuracy. We integrated the proposed A-S CB to the classic SSD, namely A-S SSD, which can be trained from scratch. Experiments on RDISD_SAR proves our approach reaches competitive performance in both accuracy and detection speed compared with the typical DSOD, and in contrast with SSD, it holds some superiorities in aspects of accuracy, speed, number of parameters, and computations.