Ensemble learning-based approach for crack detection using CNN

Crack detection is of pivotal importance in civil engineering and other related applications. Traditional methods of human inspection are tedious and severely limited. Automated crack detection by conventional image processing techniques is challenging due to their inability to discriminate crack features from background noise. Inhomogeneous lighting, shadows, and surface finish hinder the performance of digital image processing methods. The use of convolutional neural networks has helped achieve remarkably better results in the field of computer vision. Ensemble learning is an approach to aggregate the results of a number of individual models for classification or regression. Ensemble learning for crack detection has been implemented using deep convolutional neural networks (DCNN) in this paper. The models are evaluated on a number of performance metrics, namely-(i) accuracy, (ii) precision, (iii) recall (iv) Matthews correlation coefficient (MCC), (v) AUROC, and (vi) F1 score. Experimental results show the robustness of the ensembling method and offer promising scope in crack detection. They outperform the current best performance on open source concrete crack dataset. The ensemble models achieved much better performance than their individual counterparts with the best ensemble achieving a validation accuracy of 99.67%.