Study of Semantic Segmentation Models for the Detection of Pavement Degradation Using Deep Convolutional Neural Networks

The rising popularity of deep learning in pavement degradations detection stems from its remarkable benefits in capturing intricate features and addressing nonlinear problem modeling. The progress in deep learning technologies has notably contributed to the advancement of semantic segmentation, particularly in image segmentation tasks. This has proven advantageous in the realm of infrastructure maintenance, playing a crucial role in both practical applications and research endeavors. The extraction of pavement degradations identification is a notable outcome in this context. This paper explores the utilization of U-NET for semantic segmentation, a convolutional network architecture (CNN) aimed at tackling the challenge of detecting pavement degradations in images. The investigation encompasses three distinct models: U-NET with a five layer CNN Encoder, U-NET with VGG Encoder, U-NET with Alexnet Encoder. Through experimentation, we showcase the efficiency of these models in precisely identifying pavement distress within a given dataset. Our findings reveal that, although each model possesses distinct strengths, the VGG-based transfer learning model outperforms others in terms of precision and recall. This study not only adds to the expanding pool of insights into infrastructure maintenance through deep learning but also enhances data quality within the realm of deep learning-based infrastructure maintenance. Furthermore, it furnishes practical insights for professionals aiming to employ automated systems for pavement inspection.