Semantic Segmentation of FOD Using an Improved Deeplab V3+ Model

Foreign Object Debris (FOD) can cause security risks in the process of airplane take-off and landing, so the detection of FOD on the airport runway is a critical responsibility. In this article, we propose a multi-category semantic segmentation of FOD based on CBAM-Deeplab V3+. The dataset adopted in this study is a three-channel dataset obtained by an 3D laser scanning camera. The target region of image segmentation is manually calibrated on the grayscale images. Firstly, a soft attention mechanism is applied to the grayscale images as weight vectors after preprocessing the depth images. Secondly, we use the Resnet101 residual network as the backbone network of the CBAM-Deeplab V3+, with a convolutional block attention module (CBAM) applied in the spatial pyramid pooling part (ASPP) to highlight features. Thirdly, the loss function is calculated and the model is optimized using the focal loss function with dynamic weight. In this study, the segmentation effect of the model is evaluated by four kinds of common foreign objects in the airport, including wood products, screws and nuts, stones, pliers. Experiment results show that our method achieves 93.42% Mean Pixel Accuracy (MPA) and 73.69% Mean Intersection over Union (MIoU). The proposed method has achieved satisfactory results in the field of FOD image segmentation and can be further applied to clear the FOD objects using a mobile manipulator.