Segregation Method for Pothole and Manhole Features Segmented in Pavement Smartphone Images Through Deep Learning

Pavement defect segmentation is challenged by the presence of non-defect road features such as manholes that are often wrongfully detected and segmented as potholes. Incorrect detection of potholes leads to over or underestimation of intervention needs. Overestimation results in budgetary strains, whereas under estimation compromises user safety. Additionally, wrong pothole detection impacts autonomous and assisted driving safety in real-time. This study develops a segregation method incorporating edge verification to deep learning to improve the accuracy of defect detection, classification, segmentation and quantification tasks to enhance pavement management decisions. The study adds a defect verification step that includes estimating mask contour edge shapes to the You Only Look Once (YOLO) algorithm. The study proposes an algorithm, YOLO-Edge, that augments deep learning with edge verification. Detected objects with consistent mask edges are classified as non-defects (manholes) and those with inconsistent mask edges as defects (potholes). The applicability of the model is tested through experimentation on smartphone images obtained from the publicly available Road Damage Dataset 2020 (RDD-2020).