Pothole Detection of Road Pavement by Modified MobileNetV2 for Transfer Learning

Potholes are a type of pavement problem, and locating potholes is important. Asphalt road repair and maintenance are essential for public safety. When planning preventive repairs for an infrastructure such as roads, it is crucial to first identify how good is the condition of the pavement. The damaged road has potholes, cracks, lanes, and shading. In this study, we have studied pothole-detecting methods and created an accurate modified MobileNetV2 (MMNV2) methodology. This research proposes a modified MobileNetV2 model for feature extraction, image classification, and detection utilizing deep learning (DL) by implementing the transfer learning technique. Added five different layers to a pre-trained MNV2 model to increase the model performance and classification accuracy for normal and pothole image identification. This approach developed a Python model trained and evaluated on 1,300 pavement images. On combining transfer learning with deep neural network (DNN) design, this study’s results showed our MMNV2 method proved effective. This proposed modified MobileNetV2 model obtained 99.23% accuracy and 0.77% error rate with lesser parameters than previous models.