Integrating Transfer Learning and U-Net for Accurate Detection and Segmentation of Asphalt Pavement Bleeding

Abstract

Bleeding is a serious defect in asphalt pavement because it decreases vehicle traction and negatively affects the visibility of drivers. If not detected in good time, bleeding may cause serious traffic accidents in inclement weather conditions and lead to more severe damage to pavement, such as picking up of the seal coat. This paper leverages the capability of deep learning approaches for detecting bleeding in asphalt pavement. The presented method includes a deep transfer learning-based model for block-wise detection and a deep semantic segmentation for pixel-wise classification. Capable convolutional neural networks (CNNs), including the Alex network, Visual Geometry Group network, Google network (Inception), mobile network, and residual network (ResNet), are used for detecting the surface distress on the pavement surface. Based on the classification results of the CNNs, the deep semantic segmentation model of U-Net is employed to perform pattern recognition at the pixel level. To train and verify the proposed approach, an image dataset has been established from a field survey in Da Nang (Vietnam). The structure of the image classification and semantic segmentation models are trained by the adaptive moment estimation algorithm. Experimental results show that ResNet with 50 layers (ResNet50) achieves outstanding classification accuracy at detecting bleeding occurrences with a classification accuracy rate of 99.3% ± 0.5% and an F1 score of 99.3% ± 0.5%. In addition, U-Net demonstrates robust segmentation performance with an intersection over union of 0.79 and an F1 score of 0.88, effectively localizing bleeding areas with precision.