This study presents a multi-stage transfer learning approach for improving traffic sign recognition performance under both normal and low-light conditions, addressing the gap between existing datasets and the real-world road environments of the Philippines, where poor lighting, faded signs, and unstructured roads are common. A curated local dataset of 7 commonly encountered traffic sign classes comprising approximately 5,000 manually localized images was constructed and split into training, validation, and test sets (70–10–20 ratio). Five model configurations were developed and compared: a VGG-inspired baseline trained from scratch, a standard ResNet50 transfer learning model, a multiphase ResNet50 model pretrained on the GTSRB dataset, and two corresponding variants enhanced using Zero-DCE low-light preprocessing. The baseline achieved 92.17% accuracy, while the standard ResNet50 models performed similarly with and without Zero-DCE (92.10–92.45%). The multiphase ResNet50 significantly improved accuracy to 96.43% by leveraging domain-aligned pretraining, and the highest performance was achieved by its Zero-DCE-enhanced counterpart at 98.21%, showing more balanced metrics and improved recognition stability. These results indicate that low-light enhancement alone does not guarantee better performance, but becomes highly effective when paired with a feature extractor already specialized in traffic sign features. Overall, the proposed multiphase, Zero-DCE–assisted pipeline provides a strong and scalable solution for traffic sign recognition in low-visibility Philippine conditions, with potential applications in ADAS and autonomous driving systems.
Tomas et al. (Thu,) studied this question.