Deep Convolution Neural Network-Based Crack Feature Extraction, Detection and Quantification

Timely crack detection and evaluation is essential to prevent the further deterioration of the structural damage, and convolutional neural networks (CNNs) have potential applications due to their powerful feature extraction capability. Therefore, this paper employs some well-known CNN models to detect cracks and reveal the extraction features of the CNN. Subsequently, a state-of-the-art pixel-level segmentation CNN (DeepLab_v3+) was employed for crack segmentation, and the physical properties (the length and width, and so on) of the cracks were calculated according to the segmentation results. The results confirm that the CNN can effectively extract the features of different categories of cracks, and the transferred SqueezeNet has the best performance in the classification of the cracks, with competitive computational speed and accuracy, and the CNN trained on concrete cracks can also effectively detect cracks in asphalt pavement and walls. DeepLab_v3+ can also achieve the ideal performance of cracks segmentation (the MIoU, accuracy and F-score were 0.80, 97.5 and 0.78%, respectively). Finally, the proposed quantization method based on MATLAB platform achieves the desired accuracy, and the average relative errors of the length, average width, maximum width, area and cracking ratio were 7, 2, 22, 27 and 27%, respectively.