LE-YOLO: Lightweight and Efficient Algorithm for Aerial Small Target Detection

Aiming at the problems of limited edge equipment resources and poor effectiveness of small target recognition in aerial images, a lightweight and efficient algorithm LE-YOLO is proposed based on the baseline model YOLO11s. Firstly, a minimal target detection head is added to improve the detection precision by using the P2 layer feature with higher resolution. Then, a lightweight asymmetric detection head (LA-Head) is designed, and depthwise separable convolution is used to reduce the amount of model parameters and calculations. Finally, a lightweight and faster feature extraction architecture (LF - Bottleneck) is designed, which combines gating mechanism and partial convolution to design bottleneck architecture to further reduce the number of model parameters and computation. Experiments on the public dataset VisDrone 2019 show that compared with YOLO11s, mAP50 is increased by 2.1% and mAP50-95 is increased by 1.3% under the premise that parameters and calculations are reduced by 24.9% and 4.9% respectively.