Pothole Detection Using Transfer Learning Models: A Comparative Study

In this fast-paced modern world potholes are considered as some random holes on the surface of the roads and are considered as mere obstacles while traveling. But reality is much harsher than these considerations as these mere potholes are solely responsible for a significant amount of road accidents which involve hundreds of deaths and much higher property damages. This is a detailed comparative study of some popular deep-learning algorithms. The main objective of this comparative study is to find a better solution to tackle the pothole problem faced by the countries whose economy is based mostly on transport systems. The base model of these algorithms is tweaked to bring out their best results on the used dataset. Results are decided based on the output accuracy delivered by the respective algorithms. These algorithms include CNN, VGG19, VGG16, InceptionResNetV2, InceptionV3, MobileNetV2, and Xception. The MobileNetV2 with layer freezing has emerged as the best of all models used in this study with an accuracy of 96.37%. It has also taken the least computational time for each image.

The dataset that was used for this research was taken from Kaggle as it is the largest worldwide data science community, providing powerful tools and useful resources to help you achieve your data science goals. The data used contains images of various conditions of roads, including rainy environments, waterlogged potholes, camouflaged potholes, and many more, which satisfies a future work from the literature review that needed detection of potholes in extreme conditions. The potholes were also captured from different ranges, from a very close range to a far range which will make the model more versatile for detection. Another condition of taking pictures from the dashboard camera of the cars was also satisfied as the used dataset contains images taken from inside the cars. The data type of all images is in JPEG format as it has reduced file size, faster data loading, less memory usage, and less bandwidth usage. The dataset is divided into 2 parts: train part and test part, containing a total of 6096 images. The training part is also further divided into 2 parts one contains pothole images another contains plain road images; this is the same with the testing part. The data is divided by 80% to the training and 20% to the testing, allocating 5075 images in training and 1021 images in testing. Training contains 2508 plain road images and 2567 pothole images, on the other hand in testing there are 509 plain images and 512 pothole images.

A pothole is generally a hole formed on a road by erosion. Depending on the extent of the damage, their sizes vary from small to large. Their increased sizes also increase the damage it does. Early detection can decrease the amount of these damages. To address this problem, various algorithms have been employed, including CNN, VGG19, VGG16, MobileNetV2, InceptionResNetV2, InceptionV3, MobileNetV2, and Xception. Among these models, MobileNetV2 emerges as the best performer, achieving a 96.37% accuracy rate. It also has the best precision, recall, f1-score, and computational time acquiring 96.44%, 96.38%, 96.37%, and 193ms respectively. Considering the accuracy achieved, overall performance, and the computation time it takes for each step, MobileNetV2 is the best choice for Pothole detection.