Airport runway FOD detection method based on convolutional neural network

Abstract

The invention provides an FOD detection method based on a convolutional neural network (CNN). According to the method, a target candidate region is generated for an input image mainly based on a Faster R-CNN algorithm framework, and meanwhile, DenseNet is adopted to replace traditional VGG16-Net for feature extraction, so that network parameters can be greatly reduced, target features can be fullyutilized, and detection of small-size FOD is facilitated. In addition, a loss function of classification in an RPN layer is also improved, and Focal Loss is used for optimizing weights of positive and negative samples, so that a training result is focused on a small-size FOD target which is difficult to classify in the samples. Experiments show that the method has a good real-time detection effect, high detection accuracy and good anti-interference performance. The airport runway FOD image data set mainly comprises four types of objects (small steel balls, metal nuts, large screws and small screws). Compared with a classic Faster R-CNN, the method has the advantages that the FOD target detection accuracy rate is 93.93%, the FOD target detection accuracy rate is improved by 14.91%, and thedetection speed is also improved by more than one time.

Description

technical field [0001] The invention relates to the technical field of image processing and target recognition, in particular to a method for detecting Foreign Object Debris (FOD) on an airport runway based on a convolutional neural network (Convolutional Neural Networks, CNN). Background technique [0002] Airport FOD has the potential to cause damage to aircraft such as scattered aircraft parts, metal tools, concrete blocks, luggage components, wildlife and vegetation, etc. The airport runway is mainly the activity area of ​​the aircraft, and human activities are less involved. Therefore, most FODs are dropped from aircraft parts, mainly metal materials, including large and small screws, small nuts and steel balls scattered on the runway. Because the engine will generate strong suction when the aircraft tires roll at high speed, especially during the take-off and landing phases, the aircraft itself is relatively weak in response to FOD, which will cause serious damage to ...

AI Technical Summary

Problems solved by technology

These traditional image recognition methods can achieve better detection results for large-size FOD on airport runways, but the detection performance for small-size FOD will be greatly reduced.

Images