Research On Airport Runway FOD Detection Algorithm Based On Transfer Learning And Attribute Enhancement

Foreign Object Debris(FOD)will not only cause huge economic losses to airport equipment,but also threaten the life safety of airport personnel.With more and more attention paid to aviation safety,FOD detection of airport runway has come to be an crucial research topic.Research on FOD detection algorithm is an crucial step to promote the intellectualization of FOD automatic detection system.At present,domestic research on FOD detection algorithm is relatively few.Based on the existing research results in the field of computer vision,this thesis studies a more effective FOD detection algorithm for airport runway according to the characteristics of FOD.The main contents of this thesis are summarized as follows:1.Aiming at the problem of lack of open FOD detection data set of airport runway at present,according to the common object detection data set standard and FOD detection requirements,this thesis constructs FOD detection data set applied to airport runway scene.However,there are two problems in FOD data set.First,strict requirements for airport entry and exit and limited data collection time lead to small size of FOD data set and unbalanced samples.Second,FOD data set is difficult to collect all foreign bodies that may appear on the runway of the airport,and unseen FOD may be encountered in application.2.To solve the problem of insufficient and unbalanced training samples in FOD data set,this thesis extracts the same or similar data as the FOD from the public data set as the source domain sample,and use the FOD data as the target domain sample,and proposes an airport runway FOD detection algorithm based on domain-invariant feature transfer to transfer the knowledge from the source domain to the target domain.Firstly,the feature extraction network based on parameter sharing extracts the features of source domain data and target domain data.Then feature decoupling module is used to separate domain invariant feature and domain specific feature respectively.Finally,multi-stage optimization training is carried out by using domain invariant features of source domain,domain invariant features of target domain and domain specific features to realize feature transfer.In the experimental simulation stage,different domain invariant feature transfer algorithms are obtained based on the improved Faster R-CNN,SSD,YOLOv3 and YOLOv5.The comparative experiments show that the algorithm can enhance the model’s ability to characterize the FOD features and improve the FOD detection performance.3.Aiming at the unseen FOD detection problem,this thesis proposes an airport runway FOD detection algorithm based on attribute learning and background awareness by referring to the zero-shot detection.Firstly,attention convolutional neural network is designed to enhance the focus of FOD.Then,according to the singleness of airport runway application background,the attribute labels are designed for FOD category and background,which are embedded with word vector to establish a relationship with visual features.Meanwhile,the attribute attention module is designed to guide the location of discriminative attribute regions.Finally,visual features and embedded attribute information were combined to learn the cosine metric space,and the unseen category FOD detection was realized by comparing the similarity between the predicted attribute information and the embedded unseen category attribute information.Several groups of comparative experiments show that the algorithm effectively uses the target attribute information and background information,and improves the detection performance of unknown FOD.