A Compact and High-Efficiency Anchor-Free Network Based on Contour Key Points for SAR Ship Detection
Fei Gao, Changxin Cai, Wentao Tang, Yiming He
Abstract
As computer vision advances, synthetic aperture radar (SAR) ship detection increasingly incorporates deep learning techniques based on convolutional neural networks (CNNs). Presently, these approaches predominantly depend on large network architectures and anchor-based object representation methods. Although such networks can yield superior results, they are often inefficient in detection. For SAR ship detection, this letter proposes a novel network structure, the Contour Key Points Network (CPoints-Net), which utilizes a compact backbone ResNet18 and Feature Pyramid Structure (FPN) to enhance network efficiency. The Contour Key Points object representation employed in CPoints-Net is an anchor-free method that directly locates and represents ship objects, circumventing the complex post-processing associated with anchor-based techniques. Furthermore, this anchor-free approach preserves a higher degree of object feature information, and enables direct object prediction. To address the distribution of contour key points, CPoints-Net applies pointwise feature grouping (PFG) part and dynamic feature grouping (DyFG) part to refine features and employs Deformable Convolution (DCN) to achieve a more precise fit of key points to contours. On the SAR ship detection dataset (SSDD) and high-resolution SAR images dataset (HRSID), CPoints-Net achieves <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">AP</i> <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">50</sub> scores of 96.3% and 90.5%, respectively. This network has 18.64M parameters (Params), with Frames Per Second (FPS) rates of 43.8 img/s and 32.6 img/s respectively, demonstrating the optimal performance and efficiency of this network. Its source code is located at https://github.com/Daniel-tech307/CPoints_net.git.