Crack Segmentation Network using Tversky Loss Function with Variable Alpha and Beta

Cracks pixel segmentation is one of the most important factors in the structural health monitoring system. Extracting crack regions from the image is a challenging task due to the imbalanced dataset, which is one of the highly imbalanced datasets and the variation between crack and non-crack pixels is very high and varies from image to image. The imbalanced pixels directly affect the learning process of network architecture and lead to undesired network biasness. Several loss functions have been proposed to deal with the imbalanced issue including weighted cross-entropy loss, dice loss, focal loss, and Tversky loss function. In this paper, a modified Tversky loss function with a variable magnitude of penalties i.e., α and β is proposed to accurately classify crack pixels. A crack segmentation network architecture is integrated with the Tversky loss function with variable α and ß to evaluate the architecture. Experimental result shows that the modified architecture has achieved better crack and non-crack pixel accuracy as compared to FCN, SegNet, and Deeplabv3+. Furthermore, the variable α and β have performed better than the fixed values of α and β by achieving 98.93% crack pixel accuracy and 95.39% non-crack pixel accuracy.