COMPARATIVE ANALYSIS OF YOLOV8, RESNET, AND ENHANCED EFFICIENTNET-B4 + U-NET FOR AUTOMATED ROAD DAMAGE DETECTION AND SEGMENTATION

Automated road surface damage detection and segmentation is essential for modern infrastructure maintenance and smart-city applications. This paper presents a detailed comparative study of three deep-learning approaches applied to a unified pothole segmentation dataset: (1) a ResNet-based segmentation baseline, (2) YOLOv8 as a fast detection baseline, and (3) a proposed Enhanced EfficientNet-B4 encoder with a U-Net decoder (EfficientNet-U-Net). Our experiments use a curated dataset of 780 annotated images (720 train / 60 val) derived from the Pothole Image Segmentation Dataset (Farzad Nekouei). The proposed EfficientNet-U-Net achieves the best segmentation
performance with an overall accuracy of 94%, Dice coefficient of 0.74, and Intersection-over-Union
(IoU) of 0.62. We include end-to-end workflow and architecture diagrams, extensive ablations, and a
broad literature comparison (2020–2025) to contextualize our findings. The paper demonstrates that a compound-scaled encoder (EfficientNet-B4) combined with attentive skip connections and deep supervision yields superior boundary fidelity and region overlap for pothole segmentation.