Vision-based infrastructure monitoring: pavement marking visibility classification using deep neural networks

Pavement marking visibility and retroreflectivity are critical indicators of roadway safety and operational efficiency, particularly for human drivers and automated systems alike. This study presents a comprehensive deep learning framework for classifying pavement marking conditions into good, medium, and poor using a dataset of 4 306 cropped images derived from mobile roadway video. The dataset was partitioned into 70% training, 20% validation, and 10% testing subsets using stratified sampling. The pipeline integrates optical character recognition (OCR) to extract embedded metadata such as retroreflectivity levels, image preprocessing for region-of-interest isolation, and class imbalance handling via data augmentation and stratified sampling. 6 models were evaluated across standardized training, validation, and testing subsets. Performance metrics, including accuracy, precision, recall, and F1-score, were computed for model comparison. DenseNet121 and InceptionV3 emerged as top performers, achieving macro F1-scores above 0.90 and demonstrating strong generalization across class boundaries. In contrast, EfficientNetB0 underperformed, highlighting the importance of domain-specific model adaptation. This research underscores the viability of convolutional neural network (CNN)-based image classification for infrastructure condition monitoring and offers a scalable, vision-based alternative to traditional retroreflectivity measurements.