基于YOLOv7-tiny的光伏电站多种异物检测模型OA

The photovoltaic power stations in field environments are often disturbed by foreign objects of snow,weeds,bird intrusions and bird nests.The traditional detection algorithms suffer from background interference,low accuracy for small targets and low image resolution,which can not operate stably on embedded devices.Based on YOLOv7-tiny,the lightweight detection model was constructed.The Haar wavelet downsampling module was adopted to reduce the background interference,and the mixed attention module was added to enhance the detection performance of small targets,while the pixels-wise intersection over union loss function was used to solve the detection deviation caused by tilted photovoltaic panels.The experimental results show that by the optimized model,the precision of 95.8％is achieved with the recall of 95.4％,the mean average precision of 95.3％and nearly no change in model size and inference speed.After model conversion and TensorRT acceleration,the model is deployed on the Jetson TX2 platform to realize the real-time detection.