Deep Transfer Learning for Traffic Sign Recognition
Grant Rosario, Thomas Sonderman, Xingquan Zhu · 2018
In this paper, we study using deep transfer learning to utilize knowledge gained from an already developed, large dataset of traffic signs of a specific country/region, and use this knowledge to better recognize traffic signs from another country/region, under a deep learning framework. This provides the possibility of using an already established dataset from other regions to aid in the recognition of a desired target dataset, freeing users from the burden of data gathering and labeling. We propose three deep transfer learning methods, and two of them demonstrate significantly improved accuracy compared to the simple deep learning classifier. This research shows trans ferring knowledge between deep learning classifiers can provide higher accuracy for traffic sign recognition than a model which implements only deep learning to recognize traffic signs.