UAV-Based Small Object Detection Algorithm for PV Panel Identification

With the development of society and the rapid expansion of the scale of the photovoltaic power generation industry, the UAV stereoscopic and high range operation has effectively solved the problems of wide distribution of photovoltaic power plants and various types of applications, but the problem of small and dense objects in UAV aerial images will still be a great challenge for the object detection task of UAV images. Traditional convolutional neural network (CNN) methods struggle to detect common features and handle limited information, thus failing to quickly and accurately identify small target images. To this end, we propose a brilliant fusion weight network (BFWN) model based on the YOLOv5s model. The adaptive feature fusion module uses the spatial attention facility (SAF) module to change the aspect ratio dimension of the information at different scales, then add the feature galore module (FGM) before the neck structure, and finally the C3SK module combining SK attention mechanism and C3 is used to replace the C3 module in the backbone network to achieve the ability to improve the extraction of small object feature information. We evaluated our model on the remotely sensed image dataset HRRSD and our constructed Solar PV panel dataset (SP-DT). The experimental results show that the average accuracy of the BFWN-YOLO network model ([email protected]) is improved by 2.5% and 3.2%, respectively, compared with YOLOv5s, in which the small object detection performance is improved by 11.4% and 10.7%, thus, our proposed BFWN-YOLO model can effectively improve the accuracy of small object detection in photovoltaic panel datasets and remote sensing image datasets.