Yolo and RetinaNet Ensemble Transfer Learning Detector: Application in Pavement Distress

The significance of using roads for transportation has always been important and it plays a major role in the economy of a country. However, with the increase in population and urbanization, the number of vehicles on the roads as well as the length of roads have significantly increased. This has led to issues such as cracks and potholes caused by heavy rainfall and road construction materials which pose a serious risk to road users. Therefore, it is crucial to detect and maintain these defects. To address this issue, we developed a new method that uses ensemble transfer learning to identify road damages automatically. To improve the dataset, we increased the number of images and balanced the classes by augmenting the dataset resulting in 263,360 images across eight different categories. In addition, we improved the picture quality using various image processing techniques such as sharpening, histogram equalization, grey scaling, and smoothening. We trained and validated two pre-trained deep learning models over the dataset and combined them using an ensemble approach to create a final model. Our proposed model achieved an F1 score of 0.927, which suggests that it could serve as a benchmark for road damage detection.