Vol.23, No.1, 2023, pp. 31–37
UDC:

Civil structure crack detection solutions based on Artificial Intelligence (AI) have recently emerged as powerful tools with applications in complex structures where human detection considerably fails to detect the defect. The process of crack detection involves the collection of images from a wide range of infrastructure systems with diverse characteristics such as textures, conditions, and surface appearances. Considerable robustness of these processes is essential in defect detection models and need to be sufficiently robust to address these varied characteristics. In this paper, we propose civil structure crack detection using image classification in Deep Learning (DL) framework. The proposed framework uses the representational power of the convolutional layers of Convolutional Neural Network (CNN), which essentially selects appropriate features, and hence, eliminates the need for the complex feature extraction step. Additionally, good crack detection accuracy is obtained by the proposed framework at a significantly lower execution time. The proposed model consists of a total of 36 layers with 10 convolutional and transition layers connected. ReLU activation and Batch Normalization alongside dropout layers are added for better optimisation of the model. The image dataset consists of 40,000 images with 227×227 pixels with RGB channels. Images are auto resized to 28×28 before the training process in order to fit into the input layer. The image dataset is divided into 80 % for training and 20 % for testing. The proposed model achieved accuracy in the range of 88.21 to 98.60 %. We propose the hardware development of the model using Raspberry Pi™.

Keywords: convolutional neural network, crack detection, crack classification, deep learning, ReLU activation