Exploring Explainable AI for Symbol Detection in Construction Drawings

Having information in a digitally accessible format can significantly improve the efficiency of processes by enabling automation. In the construction industry, however, much building-related information is only available in analog form, mainly paper-based drawings, which hinders the introduction of digital processes. Artificial intelligence offers a practical approach for converting these analog construction drawings into digital information. Neural networks, such as the YOLO object detection model, can efficiently identify and extract crucial information from the drawings on a large scale. However, these models operate as "black boxes," making it challenging to understand the basis of their decisions. To address this problem, this study applies techniques from explainable artificial intelligence to gain insight into the inner workings of a neural network. In particular, YOLOv8 is trained to detect symbols in pixel-based drawings, and the Eigen-CAM visualization method is utilized to shed light on the network’s internal decision-making processes. With the acquired knowledge about the inner decisions of the trained network during symbol detection, the training data set is refined to enhance the model’s overall performance, leading to an improvement in the detection of mAP50 10 points.