DEPDet: A Cross-Spatial Multiscale Lightweight Network for Ship Detection of SAR Images in Complex Scenes

Abstract

Nowadays, the intricate nature of synthetic aperture radar (SAR) ship scenes, coupled with the presence of multiscale targets, poses a significant challenge in detection accuracy. Furthermore, to reduce the financial outlay on hardware, there is also a considerable challenge in lightweighting the model. In order to resolve the aforementioned concerns, we propose a cross-spatial multiscale lightweight network, designated as DEPDet. First, a new efficient multiscale detection backbone network DEMNet is redesigned. To improve the feature extraction capability of the network, a cross-spatial multiscale convolution (CSMSConv) is designed and a new CSMSConv module CSMSC2F is constructed. Meanwhile, we introduce an efficient multiscale attention module. DEMNet is capable of more effectively extracting information pertaining to multiscale ships. Moreover, to enhance the fusion of features at diverse scales, we design a new path aggregation feature pyramid network DEPAFPN, which combines deformable convolution and CSMSC2F. Finally, we introduce partial convolution to construct a lightweight detection head module PCHead, which can be employed to extract spatial features with greater efficiency. The publicly available SAR ship datasets, SAR Ship Detection Dataset and High-Resolution SAR Image Dataset, are employed for the purpose of conducting experiments. The mean average precision (mAP) obtained was 98.2% (+1.4%) and 91.6% (+1.6%), respectively. The F1 obtained 0.950 (+1.7%) and 0.871 (+1.4%), respectively. Concurrently, the Params decreased from 3.2M to 2.1M, a decrease of approximately 34%. The floating-point operations (FLOPs) decreased from 8.7G to 4.5G, a decrease of approximately 48%. The experimental results indicate that the network achieves an effective balance between detection accuracy and lightweight effect with good generalization and extensibility.