Improved Small Object Detection Based on YOLOv9 Algorithm for Remote Sensing Images

Detecting small objects in remote sensing imagery poses significant challenges, such as background noise and limited detail, which can undermine the performance of traditional detection algorithms. This study examines YOLOv8 and its improved version, YOLOv9, which incorporates several advancements specifically designed to enhance small object detection. Key innovations in YOLOv9 include a contextual attention module that emphasizes important regions based on object context, a multi-scale feature fusion approach that merges information across different layers of the network, and a dynamic anchor box optimization that adjusts anchor sizes to suit the dataset’s characteristics. Evaluations were performed using the EORSSD, ORS-4199, and ORSSD datasets, which collectively contain around 3000 diverse images ideal for this task. Results show that YOLOv9 outperforms YOLOv8 on key metrics, including mean Average Precision (mAP), model complexity, and Frames Per Second (FPS), demonstrating the impact of these enhancements in advancing small object detection for remote sensing and surveillance applications.