Image-based Road Pothole Detection using Deep Learning Model

Road pothole detection is essential to ensure any engineering structures' health. Manual pothole detection and classification is very human-intensive work. Several sensor-based techniques, laser imaging approaches, and image processing techniques have been deployed to less the intervention of humans in road inspections. Still, these approaches have some limitations, such as high cost, less accuracy, and risk during detection, as Machine learning-based approaches require manual feature extraction for the prediction. Therefore, this proposed work aims to use deep learning modes for better pothole detection results. Several pothole datasets are available online, and deep learning-based methods require lots of data for the training; therefore, pothole images are collected from the different datasets and combined into one dataset to train the model. Augmentation is also applied to the dataset for better training, as augmentation provides images with different angles, and by fine-tuning the model consequently, records with about 98 % accuracy.