Investigation of Aircraft Surface Defects Detection Based on YOLO Neural Network

The YOLO neural network has been selected to do automatic recognition of corrosion and fatigue defects on aircraft surface in this paper. The aircraft surface defect detection models based on the YOLO neural network and the Faster-RCNN are established and compared with each other under the same condition. The experimental models of the SSD and the Mask-RCNN are introduced to analyze the accuracy of aircraft surface defect sample recognition. The training, testing, and verifying data sets of neural networks including 5 type of defect images such as skin crack, skin peeling, thread corrosion, skin deformation, and skin tear, are obtained by taking pictures and are labeled by Linux-mark tool. The experimental results show that the prediction recognition accuracy of the YOLO neural network for skin crack, skin peeling, skin deformation and skin tear are 72%, 84%, 68% and 70% respectively, higher than those of the Faster-RCNN. Compared with the SSD and the Mask-RCNN, the YOLO neural network is superior in the prediction recognition accuracy for skin crack, skin crack and skin peeling.