Reviewing the Damage Rating Index (DRI) Concept Towards its Automation

Automating the detection and segmentation of objects in images has been the basis of many self-driven protocols. Point-count stereomicroscopy such as the data collection procedure to calculate the damage rating index (DRI) follows such protocol in which objects (features) are detected, segmented, and counted. This work therefore presents the beginnings of a novel automated DRI while evaluating the variability and subjectivity of the method. The Mask R-CNN model trained on 110 annotated images taken from a severely damage laboratory made concrete specimen affected by alkali-silica reaction (ASR) and produced a model loss of 0.8, targeting 0.1 in this study. Further evaluation to graphically represent variability has shown that a cumulative DRI number converges towards the expected value as the sample size, in terms of number of analyzed squares, increases. Moreover, the observed occurrences and frequencies of the counted features were plotted as distributions which helps to reduce subjectivity in the result interpretation.