Development of a YOLO-Based Model for Bird Droppings Occlusion Detection on Photovoltaic

The obstruction of bird droppings in photovoltaic power plants can cause phenomena such as hotspots on the solar panels and a decrease in power generation. To address the problem of slow detection rates and high false detection rates caused by bird droppings, which are often small targets, a bird droppings detection method called YOLOv8-NCS (YOLOv8-based method for bird droppings detection in photovoltaic arrays) is proposed. In order to achieve fast detection and localization of small targets in complex scenes, this algorithm first introduces the RFBNet feature extraction network on the basis of YOLOv8 to achieve simple and fast multi-scale feature fusion, thereby enhancing the detection ability of small targets. Secondly, to reduce the detection load of the unmanned aerial vehicle (UAV) detection system, the WIoU loss function is adopted. Finally, the NAM (non-local attention module) attention mechanism is introduced in the C2f module to better capture the dependencies between sequences and reduce background interference, thereby improving the ability to resist background interference. Experimental results show that compared to YOLOv5, YOLOv7, SSD, Faster-rcnn and CenterNet, the proposed YOLOv8-NCS achieves higher accuracy and detection speed on the bird droppings dataset in photovoltaic arrays, with precision and recall reaching 93.18% and 91.71% respectively.