A lightweight target detection method based on partial convolution and haar downsampling

With the continuous development of computer vision, more and more target detection algorithms based on deep learning are applied to UAVs. Many researches focus on solving the problem of small target detection in aerial images, but ignore the fact that there are many limitations such as arithmetic power and load on UAV platforms. These limitations will lead to a decrease in detection efficiency and fail to meet the requirements of real-time detection. In this paper, we propose a target detector for UAVs, APH-YOLO, to balance the accuracy and efficiency of the detection process. APH-YOLO consists of two lightweight modules: the Attention Partial convolution Module and the Haar Downsampling Module. APconv enhances the spatial perception capability of the model for complex scenes; Hr_down maintains the basic structure of the image and retains key features during downsampling. We conducted extensive experiments using APH-YOLO on the datasets Visdrone. The experimental results show that APH-YOLO makes the model more lightweight and improves the detection efficiency while ensuring the detection accuracy.