Lightweight Target Detection Algorithm for Aerial Images

Abstract

Target detection for aerial images has been the focus of research. However, there are several difficulties in aerial image detection, such as complex backgrounds, high resolution, and a large number of small targets in UAV (Unmanned Aerial Vehicles) aerial images. In order to achieve high-precision detection of ground targets, a lightweight aerial image target detection algorithm LATD-YOLO (Lightweight Aerial Image Target Detector) based on YOLOv7 is proposed in this paper. Firstly, a new network structure dedicated to small target detection is proposed, and the feature extraction network and feature fusion network architectures are both lightweight. The fusion relationship between shallow and deep features is reconstructed. Then, the ELAN-OD module is proposed to reduce the model computation and strengthen the feature extraction ability of the network. In addition, the hybrid attention mechanism is added to the structure of the feature extraction network, where the effective information is extracted to enhance the learning ability. Finally, a new anchor frame position metric is introduced to improve the model’s ability to handle small targets. The experimental results show that LATD-YOLO can effectively improve the detection effect. The detection accuracy is improved by 359.6% on the ITCVD and 3.6% on the VisDrone2019 datasets respectively; the volume of the model is decreased by 67.74%; the amount of parameters is reduced by 67.91%; the amount of computation is decreased by 30.92%. Therefore, LATD-YOLO can achieve excellent detection performance in high-precision and lightweight computation scenarios.