Enhancing efficiency in evaluating alkali-silica reaction (ASR) damage: an automated approach to point-count microscopy

The damage rating index (DRI) is a valuable microscopy tool for collecting and counting data on different types of concrete cracks, such as those associated with alkali-silica reaction (ASR) induced deterioration. Yet, the procedure presents drawbacks such as time consumption and variability linked to operator experience, which has sparked debates about the subjectivity of its outcomes. Embracing the forefront of technological advancements, this study explores the practicality of automating the DRI's data collection through artificial intelligence (AI) and machine learning. Like many image processing and analysis applications that use AI, the DRI is an object classification and segmentation task. This study represents a step forward in leveraging automation to enhance the objectivity and efficiency of ASR damage characterization in concrete through point-count microscopy, along with proposing a set of tools to evaluate the outcomes from the application’s perspective for more efficient training data selection. Results show that despite obtaining acceptable performance individually, where the detector-classifier performance was found to have an accuracy of 0.744, and the crack counter accuracy was 0.988, the current version of the proposed machine still displays high variability in detecting, classifying, and counting distinct crack types. Overall, the machine overestimates ASR-induced damage, which was further verified through the Chi-square goodness of fit test, indicating that further training and enhancement of the proposed machine are required.