Traffic Sign Recognition with Transfer Learning

Traffic signs are characterized by a wide variability in their visual appearance in real-world environments. Supervised algorithms have achieved superior results on German Traffic Sign Recognition Bench-mark (GTSRB) database. However, these models cannot transfer knowledge across domains, e.g. transfer knowledge learned from Synthetic Signs database to recognize the traffic signs in GTSRB database. Through Synthetic Signs database shares exactly the same class label with GTSRB, the data distribution between them are divergent. Such task is called transfer learning, that is a basic ability for human being but a challenge problem for machines. In order to make these algorithms have ability to transfer knowledge between domains, we propose a variant of Generalized Auto-Encoder (GAE) in this paper. Traditional transfer learning algorithms, e.g. Stacked Autoencoder(SA), usually attempt to reconstruct target data from source data or man-made corrupted data. In contrast, we assume the source and target data are two different corrupted versions of a domain-invariant data. And there is a latent subspace that can reconstruct the domain-invariant data as well as preserve the local manifold of it. Therefore, the domain-invariant data can be obtained not only by de-noising from the nearest source and target data but also by reconstructing from the latent subspace. In order to make the statistical and geometric property preserved simultaneously, we additionally propose a Local Coordinate Coding (LCC)-based relational function to construct the deep nonlinear architecture. The experimental results on several benchmark datasets demonstrate the effectiveness of our proposed approach in comparison with several traditional methods.