Crack Detection in Historical Structures Based on Convolutional Neural Network

: Regular inspection and maintenance work is required to ensure the structural integrity of historicstructures, especially the masonry structures which are deteriorating due to ageing and man-made activities.The structures are typically examined by visual inspection, which is a costly and laborious procedure, and often,the inspection results are subjective. In this study, an automatic image-based crack detection system usingConvolutional Neural Network (CNN) for masonry structures is proposed to aid the inspection procedure.Previous crack detection systems generally involve handcrafted features, which are then classified byclassification algorithms. This approach relies heavily on feature extraction stage, which may not offer accurateresults as some hidden features may not be extracted. In this paper, the feature extraction process is done byCNN from RGB images, and then the softmax layer is replaced by other classifiers to improve classificationaccuracy. Three classifiers are studied, namely the CNN itself, Support Vector Machines (SVM) and RandomForest (RF). A dataset containing images of cracks from masonry structures was created using a digital cameraand an unmanned aerial vehicle. The images were used in training and validating the proposed system. Thecollected images were also used to build a 3D model using the technique based on Structure from Motion(SFM), which allowed images containing cracks to be located in 3D world coordinates. It was found that thecombined CNN and SVM model performs the best among other methods with the detection accuracy ofapproximately 86% in the validation stages and 74% in the testing stage. As shown in this paper, the integrationof CNN and other classifiers can improve detection accuracy. In addition, it was shown that the system can beused to detect cracks automatically for the images of masonry structures, which is useful for the inspection ofheritage structures.