HYBRID ALGORITHM FOR NON-STATIONARY FAN ORIENTATION DETECTION BASED ON IMPROVED YOLOV5

In the context of computer vision being widely used in various industries, in order to realize the intelligent inspection in wind farm using unmanned aerial vehicle, a hybrid detection algorithm based on improved YOLOv5 is proposed. The α-CIoU was introduced by replacing the original IoU loss function to improve the precision of detection of high IoU object. The global attention module was added to the original network to enhance the global extraction ability of the network for target position and channel features. Features from accelerated segment test (FAST) algorithm were combined after the detection layer to capture the geometric features of different orientations of the fan blade to assist in fine-grained target detection, improving the detection accuracy of the algorithm and reducing the false detection rate. The experimental results show that the mean precision (mAP@0.5∶0.95) of the hybrid algorithm can reach 84.01%, which is 9.56 percentage points higher than the original algorithm, but the detection speed is only reduced by 7 FPS.