SMN-YOLO: Lightweight YOLOv8-Based Model for Small Object Detection in Remote Sensing Images
Xiangyue Zheng, Jingxin Bi, Keda Li, Gang Zhang, Ping Jiang
Abstract
The remote sensing image object detection has advanced significantly; yet, small object detection remains challenging due to their limited size and varying scales. Furthermore, real-world deployment often requires algorithms optimized for fewer parameters and faster inference. To address these issues, we propose SMN-YOLO, a lightweight small object detector based on YOLOv8n. Our approach introduces spatial-channel decoupling downsampling to reduce model size while retaining crucial downsampling information. We also present lightweight and efficient feature pyramid network (LEFPN), a lightweight multiscale feature fusion network incorporating coordinate attention (CA) to capture spatial location cues, enhancing small object detection. In addition, a multiscale feature attention module (MSFAM) further strengthens feature representation. To improve accuracy, we integrate new complete intersection over union (N-CIoU) bounding box regression loss, which minimizes the impact of positional changes on IoU, helping the model focus on low-IoU objects. Experimental results on the vehicle detection in aerial imagery (VEDAI) and AI-based tiny object detection (AI-TOD) datasets show that SMN-YOLO outperforms baseline models with a 3.2% and 2.9% improvement in mean average precision (mAP) at 0.5, respectively, while significantly reducing parameters and only slightly increasing inference time. The proposed model achieves a strong balance between performance and complexity, surpassing several leading detection models.