L-SPADNet: a target detection and classification network for high-resolution synthetic aperture radar (SAR) imagery directed toward ships

Synthetic aperture radar (SAR) imagery possesses significant applications across multiple domains owing to its superior resolution. and all-weather observation capability. However, ship detection in SAR images faces challenges such as direction arbitrariness, mutual occlusion, dense alignment, and multi-scale. The common scientific problem of these challenges lies in how to effectively extract and express ship features to realize accurate target detection and classification. To address this core problem, this paper proposes a directed ship classification and detection network (L-SPADNet), which optimizes multi-scale feature processing and enhances the ability to capture and express deep features through an innovative space-to-feature computation module, separable large-size cascade convolution, and angularly-distributed focal-loss joint network. In particular, L-SPADNet improves the training for the rotating bounding box task by introducing Angle DFL (Distributed Focus Loss with Angle Factor) and realizes the accurate detection of ships with arbitrary directions in SAR collections. The experimental findings indicate that L-SPADNet demonstrates outstanding performance on the SRSDD and SSDD+ datasets. Meanwhile, the internet has been lightweight, and the model size is only 5.08MB, which is convenient for practical deployment and application.