LAUSR

In the aircraft manufacturing industry, rivet inspection is a vital task for the aircraft structure stability and aerodynamic performance. In this article, we propose a novel framework for fully automated rivet flush measurement, which is the key step in rivet inspection task. To efficiently perform rivet flush measurement, we first develop a mobile 3-D scanning system to automatically capture the 3-D point cloud of the aircraft skin surface. Subsequently, rivet regions are extracted through point cloud processing techniques. Instead of relying on handcrafted features, we propose a novel data-driven approach for rivet point extraction via a deep-learning-based technique. Our algorithm takes a scanned point cloud of the aircraft skin surface as input and produces a dense point cloud label result for each point, distinguishing as rivet point or not. To achieve this, we propose a rivet region classification network (RRCNet) that can input the 2-D representations of a point and output a binary label indicating that the point is rivet or nonrivet point. Moreover, we design a field attention unit (FAU) to assign adaptive weights to different forms of 2-D representations via the attention mechanism in convolutional neural networks. The extracted rivet regions can then be used to perform rivet flush measurement. The abovementioned components result in a fully automatic contactless measurement framework of aircraft skin rivet flush. Several experiments are performed to demonstrate the priority of the proposed RRCNet and the effectiveness of the presented rivet flush measurement framework.