Research On Airport Runway Images Recognition Technology Based On CNN

With the development of industrial technology,aviation has become one of the most popular means of travel in today's society.The airport runway is the only way for planes to take off and land.In recent years,foreign object debris(FOD) around the airport runway have led to more and more air accidents.Manual inspection of foreign object debris on the airport runway is time-consuming,laborious and difficult to avoid the false inspection caused by fatigue.Therefore,an efficient automatic identification system of foreign object debris on the airport runway is of great significance to the guarantee of aircraft landing safety.In view of the existing airport runway recognition systems which are based on radar wave and traditional vision of recognition methods,this article is based on a kind of omnidirectional mobile airport runway inspection robot system,studies the convolutional neural network(CNN) airport runway foreign object debris image recognition method.This method can use the optical camera of the robot system to obtain the image of the airport runway,and use the convolutional neural network to detect the presence of foreign object debris on the runway,and identify the types of foreign object debris.First of all,this paper compared the performance of AlexNet,VggNet-16 and GoogLeNet,three most commonly used convolutional neural network models.Otherwise Caffe and Tensor Flow,two mainstream deep learning frameworks,on the foreign object debris recognition task of the airport runway,and selected the most suitable convolutional neural network model VggNet-16 for the system.Experiments show that the network model has high training speed,high recognition accuracy,and can complete the recognition task well in the case of occlusion,fuzzy target recognition and other difficult situations,and it is easy to improve the network structure under Caffe framework.Secondly,this paper studied in depth,in the process of the establishment of the training sample library based on keras library functions and the depth of the convolution Deep Convolutional Generative Adversarial Networks(DCGAN)enhancement method,two kinds of data on the basis of the original large number of samples,effectively amplification,the sample database by using data obtained from two kinds of methods to enhance more rich diversity,improve the network generalization ability and robustness of the model.In order to solve the problem of partial effective information loss in the process of feature extraction,a convolutional neural network(CNN) based on multi-level feature fusion was used to identify the foreign object debris in the airport runway.The specific method is to fuse the feature images of the 4th convolutional layer and the 5th convolutional layer in the convolutional neural network Vgg Net-16,so as to preserve more effective information of the same image and send it into the classifier at the same time.Through experimental comparison,the recognition accuracy of the multi-feature fusion technique used in this paper is obviously better than that of the common convolutional neural network,and to a large extent better than the traditional neural network recognition algorithm.In this paper,the establishment of the above CNN model and the algorithm training method are presented,and the runway foreign object debris image recognition simulation is carried out through the established model,demonstrating the stability and superiority of the proposed method.In the end,this paper summarized the shortcomings of the runway image recognition system and the future research direction are prospected.