Real-time detection and localization of honeycomb defects in concrete pillars using hybrid deep learning models

Abstract

This paper presents a hybrid model based on deep learning for the detection and instance segmentation of defects in honeycombs of concrete structures with YOLOv5 and Mask R-CNN. The approach combines the fast object detection feature of YOLOv5 with the precise instance segmentation feature of Mask R-CNN to effectively resolve and localize defect areas in structural images. A silicon dataset containing 1991 annotated images was utilized to train the model and evaluate. The system contains upgraded preprocessing, normalization, and Non-Maximum Suppression (NMS) to confirm robust and best performance. The model attained 98.26% training accuracy and 97.80% validation accuracy. Experimental results show very high efficacy over various measures, such as a Dice Similarity Coefficient of 0.9210, Matthews Correlation Coefficient of 0.9620, mean Average Precision (mAP) of 0.9752, F1-score of 0.9835, Precision of 0.9843, Recall of 0.9812, PR-AUC of 0.9752, IoU score of 0.9515, and Calibration Curve Error of 0.1800. The proposed method provides high accuracy, superior generalization, and robust segmentation performance, thus, it is highly suitable for real structural flaw inspection in construction and civil infrastructure health monitoring.