基于深度学习的蒙皮表面缺陷检测研究

Research on detection of aircraft skin defects based on deep learning

采集大量蒙皮表面缺陷图像作为训练样本,在Faster RCNN网络和SSD网络的基础上设计缺陷检测网络。制定网络评价准则,通过查全率、查准率等技术指标全面衡量网络的可用性。利用高性能图形工作站对2种网络进行训练和测试,测试结果表明,2种网络均可以检测并识别典型缺陷,标识缺陷位置、类别和置信度。结果表明:基于2种基础网络的缺陷检测网络能检测并识别典型缺陷,相关技术指标达到预期水平。

In view of the rich two-dimensional texture features of aircraft skin,such as paint drops,scratches,corrosion and holes,it is advantageous to use visible light images to detect and identify defects,collect a large number of skin surface defect images as training samples,and design a defect detection network based on Faster RCNN network and SSD network.This paper establishes the network evaluation criteria,through the recall,precision and other technical indicators to comprehensively measure the availability of the network.High performance graphic workstations are used to train and test the two networks.The test results show that the two networks can detect typical defects and identify the location,category and confidence of the defects.In terms of network evaluation results,the two networks have their own characteristics and can be used in combination with actual scenarios.The results show that the defect detection network based on the two kinds of basic networks can detect and identify typical defects,and the related technical indexes can reach the expected level.