Concrete section segmentation with advanced deep learning models and refined labeling approaches

Abstract

This study investigates deep learning models for semantic and instance segmentation of construction site images, focusing on concrete sections. A labeling scheme was developed for the image dataset, extending to multiple classes to achieve more detailed segmentation. ResNet-50, Xception, Inception ResNet V2, and Mask R-CNN models were trained and evaluated using key metrics, including accuracy, intersection over union, boundary F1 score, and receiver operating characteristic curves. The influence of hyperparameters, particularly the learning rate, was thoroughly analyzed. The results demonstrated that the optimal learning rate is 0.01, which led to a performance improvement within the studied range. Additionally, the Xception model consistently outperformed the others across most classes, delivering robust accuracy and reliability,with an accuracy value of 94%, an IoU of 88%, and a BFS of 73%. Furthermore, this study examines the impact of practical factors, including brightness, blur, camera rotation, perspective distortion, and illumination, on segmentation performance. The findings reveal that segmentation accuracy declines significantly under extreme conditions, highlighting the necessity of data augmentation and high-quality image acquisition to improve model resilience.”