Application of EfficientNet Transfer Learning with Incremental Fine-Tuning for Road Damage Detection

Abstract

Image-based road damage detection is an essential component of intelligent infrastructure monitoring systems. However, conventional transfer learning often fails to adapt pre-trained models to domain-specific characteristics such as fine-crack textures, illumination variations, and perspective distortions. This study proposes an EfficientNet-based road damage classification model that leverages incremental fine-tuning and multi-stage data augmentation to enhance feature adaptation and model robustness. The experiments were conducted using the Road Damage Detection dataset from Kaggle, comprising 1,400 labeled images across several road damage classes. The dataset was partitioned into 80:10:10 splits for training, validation, and testing, with stratification. The proposed approach gradually unfreezes EfficientNet layers through a structured incremental fine-tuning schedule while applying staged augmentation to expand data diversity. Experimental results show that the baseline EfficientNet transfer learning model achieved 78.26% accuracy, whereas the proposed model improved performance to 97.10% accuracy, with 97.60% macro precision, 97.20% macro recall, and 97.30% macro F1-score. The results demonstrate that incremental fine-tuning effectively enhances feature adaptation to road damage textures, while multi-stage augmentation improves model robustness. These findings indicate that the proposed approach provides an effective strategy for improving deep-learning-based road damage detection systems in real-world infrastructure monitoring applications.