Inshore Ship Detection in SAR Images Via an Improved SSD Model with Wavelet Decomposition

Abstract

Inshore ship detection is a meaningful but difficult application for synthetic aperture radar (SAR), due to the land targets affection and the complex environment near the coast. In order to improve the capability for ship detection in complex environment, a novel method based on wavelet decomposition and improved single shot multi-box detector (SSD) is proposed. A wavelet decomposition is performed firstly to obtain both high- and low-frequency components of SAR images. Then, the two components are used to generate new images containing the texture information of SAR images. After that, new training images are fed into residual network (ResNet) that joins the squeeze-and-excitation (SE) blocks. At last, SSD model uses the feature maps output by ResNet for detection. The effectiveness of the proposed model is testified through numerical experiments, and it is shown that the proposed model has better performance in inshore ship detection using SAR images compared with the traditional SSD model.