A CUSTOMIZED NEURAL NETWORK MODEL FOR DETECTING FOREIGN OBJECT DEBRIS IN UAS-BORNE SAR IMAGES

Foreign object debris (FOD) can cause serious accidents and economic loss. Synthetic aperture radar (SAR) aboard unmanned aerial systems (UAS) can provide high-resolution images independent of light and weather conditions. In this paper, a FOD detection model is proposed. It is based on the popular neural network model known as You Only Look Once (YOLO). FODs are small objects with few pixels. For this reason, an additional head is inserted into YOLOv8 to enhance the ability to detect small objects. Moreover, sliceassisted hyper-inference (SAHI) is employed during the prediction stage to further improve the detection rate. Finally, the model was trained using real experimental data obtained from UAS-borne SAR, and tested on unseen images. The results demonstrate that the proposed model can effectively and efficiently detect FOD with a detection rate of 83.3%.