SODA-Net: A Lightweight Small Object Detection Network for Drone-Based Optical Sensor Systems

Abstract

The widespread deployment of drone-based optical sensor systems in environmental monitoring, traffic management, and security surveillance has made small object detection a critical task in aerial sensing. However, the sensor-captured objects are typically small, low-resolution, and frequently occluded, posing challenges for achieving high accuracy under resource-constrained conditions. To address these challenges, this article proposes SODA-Net, a lightweight small object detection network tailored for drone optical sensor imagery, enabling efficient real-time inference on edge devices. The proposed method integrates four key modules to enhance feature representation with low computational cost. A shallow feature enhancement (SFE) module improves early feature extraction and semantic fusion. A detail enhancement and feature fusion (DEFF) module strengthens edge and texture representation using lightweight reparameterized convolutions. An attention-guided lightweight downsampling (ALDS) module preserves sensitivity to small objects while improving feature utilization. In addition, a ghost shuffle-based spatial pyramid pooling-fast module enlarges the receptive field for global context aggregation and multiscale fusion. Experiments on the VisDrone2019 dataset show that SODA-Net improves mAP50 by 4.4% while reducing parameters by 77.3%. Further validation on TinyPerson and aerial image tiny object detection (AITOD) demonstrates strong generalization, with mAP50 gains of 5.0% and 6.8%, respectively. The results indicate its effectiveness for real-time deployment in resource-limited aerial optical sensor systems.