A Wavelet Feature Based SAR Ship Detection Algorithm in Sparse Framework

This study proposes a novel method for ship target detection in Synthetic Aperture Radar (SAR) images. Traditional ship detection methods rely heavily on manually designed features, which are less robust. In contrast, deep learning-based detection methods, which can autonomously learn parameters and extract features, have achieved promising results in SAR ship detection. However, SAR ship target detection faces challenges such as complex backgrounds, noise interference, and scale variation. These issues are particularly pronounced in near-shore scenes where background interference and dense ship arrangements lead to occlusion. Furthermore, unlike optical RGB images that have three independent channels, SAR images have a single-channel problem, making it difficult for Convolutional Neural Networks (CNNs) to extract features from limited channel information. Additionally, most mainstream object detection methods use dense anchor boxes either directly or indirectly, which introduces a large amount of background regions with no ships in sparse SAR images. To address these issues, we first apply specific data augmentation techniques to handle background interference and occlusion in SAR ship targets. Next, we use wavelet transform techniques to enrich the channel information of SAR images. Finally, we adopt a sparse framework to cater to the sparse nature of SAR ship targets. Experiments conducted on the SSDD and HRSID SAR ship datasets demonstrate that our method has significant advantages compared to other advanced detection algorithms.