GeoMask : Foreign Object Debris Instance Segmentation Using Geodesic Representations

Abstract

Foreign object debris (FOD) prevention and clearance is a major responsibility in the aviation sector. Multiple agents/systems are responsible and involved in the FOD walk both during the pre-flight and post-flight phases. These in-spections are labor-intensive and require high technical competency to identify obstructions and damages impacting the airworthiness. Data availability, complex backgrounds, harsh weather circumstances, limited visibility conditions, and additional human effort all contribute to the difficulties experienced during FOD detections. In this paper, we propose the GeoMask, a foreign object debris instance segmentation technique that uses geodesic representations to recognize and label the foreign object. The proposed algorithm uses a convolutional neural network to deform an initial contour to match the object boundary, which also exploits the graph structure of a contour. Further, the prediction of the contour is based on the centroid mass theorem and optimized using dense distance regression in geodesic spherical representation. Data availability concerns are addressed using a custom data set and by applying various data augmentation techniques to the images before training the model. Experiments indicated that the proposed approach achieves a higher rate of accuracy and could lead to enhanced airport surveillance increasing safety and efficiency.