SS-GAN based road surface crack region segmentation and forecasting

Abstract

Automatic segmentation of crack regions is the major and complex task in an intelligent transportation system. In this proposed work, we have developed a Semantic Segmentation based Generative Adversarial Network model (SS-GAN). Various conditional factors such as the number and type of vehicles traveled during the months and environmental conditions like temperature, and precipitation will be additionally fed into the proposed SS-GAN model to segment and forecast the respective months’ road crack images. In-order to forecast similar crack images of appropriate months with better performance, we have introduced new loss functions additionally. A semantic segmentation model is developed to extract the required semantic features of the crack regions which helps to segment road crack images more accurately. Finally, the proposed model concatenates the extracted features of both discriminator and semantic segmentation models to classify types of cracks with better performance. SS-GAN model is trained using the Crack ForeCast (CFC) dataset in an end-to-end and paired translation model. A comprehensive analysis of experimental results is carried out to measure the performance of the proposed SS-GAN model. Various metrics are used to evaluate the quality of semantic feature-based segmentation for the forecasted crack images. The obtained results prove that the SS-GAN model showcases better performance than other state-of-the-art models.