Comparative Analysis of YOLOv8, ResNet, and Enhanced EfficientNet-B4 + U-Net : For Automated Road Damage Detection and Segmentation

The changing infrastructure maintenance and
smart cities apps require automatic detection and identification of road surface damage. In this paper, comparative information on techniques of image recognition has been provided at length. Particularly, the Res-UNet structure is used as a baseline of segmentation, YOLOv8m is used to identify objects in a short time, and a better encoder with a U-Net decoder (Efficient Net-U-Net) is introduced. The experiments use a selected sample of 780 image annotations of the Pothole Image Segmentation Database, stratified in a rigid proportion of 70/15/15. The mentioned Efficient Net-U-Net has the highest segmentation performance with the accuracy rate of 97.77, F1-score (Dice) of 0.9107, and Intersection-over-Union (IoU) of 0.8415. Diagrams of the workflow and architecture are included, and the literature (2021–2026) on the same studies is extensive to have a perspective. The essay shows that a combination of EfficientNet-B4 encoder and compound scaling with skip connections of concurrent spatial and channel Squeeze and Excitation (scSE) and multi-scale loss yields the best boundary fidelity. This enables the model to effectively indicate anomalies and not yield false positives in locations with water puddles but still the viable edge latency is 29.8 FPS.