ALF-Net: A Lightweight and Fast Network for Small Object Detection in Airborne Remote Sensing

Small object detection in airborne remote sensing imagery is challenged by tiny target scales, complex backgrounds, and heavy computational demands, thereby limiting real-time deployment on edge devices. To address these issues, we propose ALF-Net, a lightweight and efficient detection network specifically designed for small object detection in airborne remote sensing images. ALF-Net incorporates three key innovations: (1) an Optimized Base Structure (OBS) that preserves high-resolution features to enhance small-object representation; (2) the DualConv-enhanced Cross Stage Partial bottleneck (DC-CSP) that strengthens fine-grained feature extraction without increasing parameter complexity; (3) a plug-and-play Background Separation Enhancement Module (BSEM) with the Bidirectional Gating Attention Module (BGAM) that adaptively suppresses background interference and highlights target features. Experiments on three benchmark datasets—AI-TOD, USOD, and VEDAI—demonstrate the effectiveness of ALF-Net for small object detection under resource constraints. On AI-TOD, ALF-Net achieves a$\bf {mAP}_{0.5}$of 0.484 and a$\bf {mAP}_{0.5{:}0.95}$of 0.207, while obtaining small-object$\bf {mAP}_{s}$of 0.270 (USOD) and 0.462 (VEDAI), respectively. With only 3.07 M parameters and 21.8 GFLOPs, ALF-Net attains a latency of 57.9 ms on an embedded hardware platform (NVIDIA Jetson Xavier NX), significantly outperforming advanced methods such as FFCA-YOLO in efficiency. In addition, evaluations on RS-STOD further corroborate robustness to super-tiny objects; however, on size-balanced scenes (STAR) it may trail heavier DETR/conv detectors, reflecting an explicit accuracy–efficiency trade-off. Overall, ALF-Net achieves a favorable balance between detection accuracy and efficiency, highlighting its strong potential for real-time airborne applications.