Automated detection of exposed reinforcements in concrete structures through machine learning and image segmentation.

This study aimed to detect exposed reinforcements in reinforced concrete structures using machine learning techniques. A dataset termed "pre-training" was developed using 15 training images segmented with the Trainable Weka Segmentation tool in ImageJ. Four training strategies were evaluated: general pre-training (PT), pre-training with manual adjustments (PTA), local training on 2.5% of the area (LT), and local training with manual adjustments (LTA). The PTA configuration with 1% manual adjustments achieved accuracy values up to 0.97 and recall up to 0.83. Even without manual adjustments, the model demonstrated robust performance, achieving 0.95 accuracy and 0.73 recall. These findings underscore the potential of the proposed approach to enable precise segmentation of reinforcements while minimizing the need for manual intervention.