LSODNet: A Lightweight and Efficient Detector for Small Object Detection in Remote Sensing Images

Object detection in remote sensing images is critical in various military and industrial applications. Accurately identifying small objects remains challenging due to low contrast, noise interference, and complex backgrounds. Advanced deep learning techniques have improved detection accuracy through complex architectures and feature extraction mechanisms. However, most existing methods rely on computationally expensive networks, which limits their effectiveness across diverse remote sensing scenarios. To address these issues, we propose a Lightweight Small Object Detection Network (LSODNet) for efficient small object detection in remote sensing applications. LSODNet integrates three novel plug-and-play modules: First, a Lightweight Feature Convolution (LFC) module reduces the computation cost by splitting feature channels between depthwise convolutions and deformable convolutions, striking a balance between efficiency and feature richness. Second, a Receptive Field Expansion Block (RFEB) cost-effectively enlarges the network's receptive field through multi-branch dilated convolutions, capturing crucial multi-scale contextual information for small objects without significant overhead. Third, a Hierarchical Fusion Attention Block (HFAB) employs a two-stage attention mechanism, combining spatial attention to emphasize local object information and proxy attention to encode global context to fuse multi-scale features and suppress background clutter. Experimental results on the NWPU VHR-10, SSDD, and LEVIR-Ship datasets show that LSODNet achieves competitive or superior accuracy and inference speed with only 2.2 M parameters and 6.1 G FLOPs, while significantly reducing the computational cost compared with existing methods. The source code is available at https://github.com/LSODNet/LSODNet.