High-level Feature Learning And Damage Monitoring Of Aircraft Surface Images

The aircraft has become an important transportation in daily life.However,accidents that occur from time to time are an existed threat to people’s lives and property.In addition to the fatigue damage caused by normal flight,the incentive of aircraft accidents also includes the structural corrosion caused by the damage of aircraft surfaces.At present,the research method of the damage detection of aircraft surface images is mainly focused on the hand-crafted method based on portable robots.However,it has the drawbacks of a high rate of missed detection and slow speed which not meets the requirement of real-time detection.This paper proposed a method of the aircraft surface damage detection based on deep neural networks,which realizes automatic feature extraction for the aircraft surface damage and the location and classification function of damages.As the result,the method realizes the automatic processing of a large amount of image data and thus serves as an auxiliary method to assist the airport ground staff to timely detect and maintain damages.For the problem of aircraft surface damage detection,the dataset including 32056 pieces of raw data obtained by the flight school,aerial museum shooting,and web crawling was first made.After removing poor quality images,a total of 9928 images that can be used for neural network training were obtained by data augmentations.The corresponding labels were obtained through manual calibration.This paper applied the deep learning method to the detection.First,the Faster R-CNN,YOLO,and SSD object detection models were built to study the detection and classification effects of different network structures on the aircraft surface damage dataset.Then,the network training process and experimental results were detailed.In conclusion,SSD achieved the best test results with 68.53%mAP.Furthermore,for the unbalanced sample number problem of aircraft surface damage dataset and the irregular-shape damages which are difficult to classify and locate,this paper used Focal loss as new loss to improve SSD object detection algorithm as well as introduced the rhombic prediction box and adjusted the discriminating mechanism of the prediction frame,which effectively improves the detection result of the difficult classification sample.