Target Detection Algorithm for Drone Aerial Images based on Deep Learning

In recent years, the rapid development of drones has brought tremendous changes to many fields. From refined management in agriculture and forestry to aerial surveying in urban planning, the popularity of UAVs provides efficient and accurate data support for various industries, promoting social intelligence and technological progress. Meanwhile, deep learning methods represented by computer vision provide powerful tools for research such as image processing and object recognition. The research results of HE-YOLOX-ASFF (Highly Efficient You Only Look Once eXtend with Adaptive Spatial Feature Fusion) algorithm for target detection in UAV aerial images in this article can help improve the recognition accuracy and efficiency of drone remote sensing images, and apply it to fields such as smart agriculture, urban management, and environmental monitoring. At present, detection technologies for large-scale targets have achieved good detection results in clear backgrounds. However, drone aerial images often have complex backgrounds, small target volumes, limited available features, and high target density, making them prone to occlusion. Due to the above reasons, existing target detection algorithms have low detection accuracy for weak and small targets in aerial images under complex backgrounds, and are prone to missed and false detections. Therefore, the detection method of weak aerial targets in complex backgrounds remains a challenging issue. The experimental results show that in identifying pedestrians and bicycles, the parameter values are relatively high, at 33.3 and 65.1, respectively, while the FPS (frames per second) is relatively low, at 63 and 38, respectively. This indicates that the algorithm requires more parameters to process these targets and may affect processing speed.