CONVOLUTIONAL NEURAL NETWORK-BASED SPACE DETECTION MODEL FOR STIRRUPS AND TIES

Abstract

Stirrup and tie inspection is tedious, especially in large-scale construction projects, which could bring problems in the balance of the Iron Triangle: time, cost, and quality management. This tedious inspection process could be optimized by applying recent advancements in deep-learning algorithms, specifically neural networks, in engineering. In this study, the training program and object detection technology used were MATLAB and YOLOv2, respectively. The dataset was primarily obtained from a physical model setup, comprising an equal number of images for beams and columns. Furthermore, dataset splitting was applied to the total of 2,142 images generated post-augmentation, using a 70-15-15 split for the training, validation, and test sets. After performing centroid location and distance calculations in MATLAB, the model was evaluated using multiple regression analysis and ANOVA to analyze the relationship between the dependent variables: spacing accuracy and average precision, and the independent variables: dataset size, learning rate, and image resolution. Based on the findings, a larger dataset enabled the model to generalize stirrup and tie features better; the model with a smaller learning rate converged insufficiently within the same number of epochs, and the model performed reliably even when the test images were resized. Overall, the convolutional neural network-based space detection model for stirrups and ties performed with high accuracy, with the learning rate as the significant factor affecting the model's performance.