Enhancing Road Safety: A Deep Learning Approach to Detect and Classify Road Damage

Detecting and classifying damaged regions of road images is a critical task in computer vision. It’s particularly important for self-driving cars. We used two datasets named RDD-2020 and RDD-2022 to train our models. Several earlier approaches using this dataset were unable to reliably identify multiple types of damage within a single image. To improve this, we tried five different models for both RDD-2020 and RDD-2022 datasets. These models are composed of a feature extractor or backbone network and a detection head or detection network. We utilized ResNet-50, ResNet-101, and MobileNetv3 as feature extractors, and Faster RCNN (Region-based Convolution Neural Network), SSD (Single-Shot Detection), and YOLO (You Only Look Once) as detection heads for identifying damaged areas and classifying these damaged areas. The top-performing model was YOLOv8 with ResNet-50. Other models include YOLOv7 with ResNet-50, Faster RCNN with ResNet-50, Faster RCNN with ResNet-101, and SSD with MobileNetv3. We adjust our best model using different techniques to make it even better. We found that using a specific combination of optimizer and batch size resulted in the best performance for detecting different types of road damage. We achieved a better F1 score for YOLOv8 with ResNet-50 using a batch size of 4 and Adam optimizer using the same RDD-2020 and RDD-2022 datasets. We found an F1 score of 0.62 for RDD-2022 and 0.60 for RDD-2020.