GSFFE-Net: Global Semantic Feature Fusion Enhancement Network for Multiclass Multitarget Detection in SAR Images

Multi-class multi-target detection in Synthetic Aperture Radar (SAR) imagery aims to simultaneously, rapidly, and efficiently detect and recognize multiple targets of interest within SAR images. It plays a crucial role in both military and civilian applications. However, most existing SAR target detection algorithms are designed for single target classes. When applied to multi-class detection scenarios, they often require multiple passes over the same scene using a series of single-class models, which results in poor performance, high computational costs, and long processing times. The development of algorithms capable of detecting multiple target classes simultaneously can effectively address these challenges. Nevertheless, compared to single-class detection methods, multi-class detection algorithms must account for more complex inter-class similarities and intra-class variations. Additionally, challenges such as multi scale variations caused by diverse target sizes, background clutter, and terrain interference become more pronounced in multi-class detection scenarios. In response to the above issues, this paper proposes a novel Global Semantic Feature Fusion Enhancement Network (GSFFE-Net) to address multi-class SAR target detection with enhanced global context modeling. The proposed framework integrates three key modules: 1) Wavelet Feature Recombination Module (WFRM), which decomposes SAR images into low-frequency background semantic features and high-frequency target details via 2D wavelet transform, then recombines multi-channel features to exploit prior spatial frequency information. 2) Global Semantic Attention Spatial Pyramid Pooling Fast (GSA-SPPF) module, which employs dual global pooling branches to capture global context, combined with a Multi-Scale Convolutional Attention (MSCA) mechanism to enhance cross-channel feature synergy. 3) CMUNeXt Block, inspired by medical image segmentation, which uses large-kernel depthwise separable convolutions and inverted residual structures to model long-range spatial dependencies and retain low-level features. Extensive experiments on the large-scale SARDet-100K dataset demonstrate that GSFFE-Net achieves state-of-the-art performance, reaching 89.0%, 71.8%, and 64.3% for mAP@50, mAP@75, and mAP@50:95 respectively, surpassing the base line by 5.2%, 11.9%, and 9.7%. Ablation studies confirm the contributions of the proposed modules. This work provides a robust solution for multi-class SAR target detection in complex scenarios, while achieving a high inference speed of 769 FPS with only 3.9M parameters, demonstrating strong potential for real-world deployment.