Quantifying Traffic Sign Damage Severity with a Modified Triplet Loss Network

Effective traffic sign maintenance is crucial for road safety, yet traditional manual inspections are costly and inefficient. While AI-powered image analysis can automate damage detection, simple binary classification (damaged/undamaged) is insufficient for prioritizing repairs. This paper introduces a CNN-based system that quantifies traffic sign damage severity with a numerical score, enabling data-driven maintenance. Our core contribution is a modified triplet loss function incorporating sub-sampling, which ensures that the learned feature representations capture the spectrum of damage, from minor to severe. Unlike standard triplet loss, this prevents over-clustering of similar damage levels. We use a VGG16 architecture, pre-trained on ImageNet and fine-tuned with our custom loss function. Our dataset combines real-world images from COCO, augmented with synthetic zero-damage examples and internet-sourced images of severe damage, creating clear benchmarks. Experiments on stop signs show that while our method achieves a binary classification accuracy of 76.9%, it excels in identifying damaged signs and generating scores that correlate well with visually assessed damage levels. This approach significantly improves binary classification, providing a more nuanced and actionable assessment for prioritizing traffic sign repair and enhancing road safety.