CNN-LSTM detection and short-term prediction of concrete crack severity
TL;DR
A leakage-controlled convolutional neural network long/short-term memory (LSTM) framework for static crack detection and short-term prediction of crack severity from time-ordered image sequences that takes this method as an interpretable short-horizon monitoring framework, rather than a fully calibrated crack metrology system.
Abstract
The development of concrete cracks will affect the durability, serviceability, and safety of civil infrastructure. This study proposes a leakage-controlled convolutional neural network long/short-term memory (LSTM) framework for static crack detection and short-term prediction of crack severity from time-ordered image sequences. The temporal task is defined as a next-step prediction of the crack severity index (CSI) derived from the image, rather than claiming complete physical crack geometry prediction from bounding box annotations. The CSI is calculated based on the ratio of crack-box and region-of-interest areas. The ResNet34 encoder extracts frame-wise spatial features, and a two-layer LSTM models the temporal dependency of structure-level sequences. Sequence construction, augmentation usage, split rules, optimisation settings, and evaluation metrics are selected to improve reproducibility and avoid temporal leakage. On the dataset, the framework achieves 93.0% classification accuracy, 88.3% mean average precision (mAP)@0.5, 0.76 mean intersection over union, 0.015 normalised mean squared error, and 89.2% F1 score. It also clarifies the practical interpretation of CSI growth, discusses detector-level benchmarking, domain shift and uncertainty estimation, and takes this method as an interpretable short-horizon monitoring framework, rather than a fully calibrated crack metrology system.