SDNet: a lightweight ship detection network in remote sensing images by super-resolution enhancement and detail completion

Abstract

Ship detection in vast ocean areas using remote sensing images is crucial for military surveillance and emergency operations. Traditional methods relying on ground data-processing centers suffer from poor timeliness and high communication resource consumption. In-orbit processing is gaining popularity, but developing fast and accurate ship detection methods with limited resources remains challenging. This study proposes SDNet, a lightweight network with super-resolution enhancement, and detail completion, for in-orbit detection of tiny ships. The main detector branch uses the adaptive cross-stage partial convolution (ACPC) module to form an efficient backbone. The feature pyramid network (FPN) combines with the cross-level wavelet transform multi-head attention (CWTMA) module for ship feature extraction. The absolute IoU (AIoU) loss aids in bounding box regression. The auxiliary branch uses the residual wavelet transform (RWT) module to generate a shallow super-resolution network, enhancing the main detector’s focus on the ship area. During inference, this auxiliary branch is removed to avoid additional computation. SDNet achieves 84.2% and 91.1% accuracy on Levir-Ship and HRSID datasets, which are 1.8% and 1.1% higher than DRENet. It outperforms other lightweight detectors in both accuracy and efficiency. For 512x512 input images, it has 4.86M parameters and requires 7.9G FLOPs.