Performance Comparison of Pretrained Convolutional Neural Networks on Crack Detection in Buildings

Crack detection has vital importance for structural health monitoring and inspection of buildings. The task is challenging for computer vision methods as cracks have only low-level features for detection which are easily confused with background texture, foreign objects and/ or irregularities in construction. In addition, difficulties such as inhomogeneous illumination and irregularities in construction present an obstacle for fully autonomous crack detection in the course of building inspection and monitoring. Convolutional neural networks (CNN’s) are promising frameworks for crack detection with high accuracy and precision. Furthermore, being able to adapt pretrained networks to custom tasks by means of transfer learning enables users to utilize CNN’s without the requirement of deep understanding and knowledge of algorithms. Yet, acknowledging the limitations and points to consider in the course of employing CNN’s have great importance especially in fields which the results have vital importance such as crack detection in buildings. Within the scope of this study, a multidimensional performance analysis of highly acknowledged pretrained networks with respect to the size of training dataset, depth of networks, number of epochs for training and expandability to other material types utilized in buildings is conducted. By this means, it is aimed to develop an insight for new researchers and highlight the points to consider while applying CNN’s for crack detection task.