Fissura_Net: An Enhanced Crack Detection and Prediction Network

Rapid developments in computer vision, big data and deep learning have contributed to automated fault diagnosis in many domains. Intelligent pavement damage prediction is of great significance in transportation maintenance to ensure safety. As the depths of convolution neural networks increases, loss function’s gradient move towards zero. Furthermore, severity of the crack’s inhomogeneity, intricacy of the background—which includes the low contrast with the nearby pavement and the possibility of shadows with a similar intensity results in ambiguous crack prediction. Hence, the crack detection image dataset is preprocessed using homomorphic filtering. To avoid the complications of vanishing gradient, this paper presents three different deep learning architectures (custom model (CR Net), Res-Net50 and Fisura_Net) based on residual and skip connections with hyper parameters tuned optimally to classify and identify crack and no-crack images. Experimental analysis was carried out with a comparison of contemporary CNN models like ResNet-50, ResNet 152 and previous existing research works on the basis of precision, recall, accuracy, F1 score, loss accuracy, and receiver operating characteristic (ROC) curves and Area Under Curve (AUC) metric. The comparative results ascertain the proposed Fisura_Net model utilizing less computational complexity with highest classification prediction efficiency (99.85%), elevated F1 score (1.0) and AUC score (0.99). Fisura_Net outperformed the other existing models and exhibited promising damage prediction.