A Long-Tail Regularization Method for Traffic Sign Recognition Based on Parallel Vision

Abstract

The long-tail effect is prevalent in the actual world, and the resulting long-tail problem hinders the development of intelligent vehicles towards a more reliable and safe direction. In previous studies, the theory of Long-tail Regularization (LoTR) based on parallel vision was proposed, and further research and experiments on the long-tail problem of traffic sign recognition are conducted in this paper. Specifically, in the process of long-tail regularization for traffic sign recognition, a novel generative adversarial network is designed and improved to generate images. By feeding the category-specific image from standard gallery into the proposed network, it can generate a large number of virtual images with the same category. Finally, the generated virtual images are added to the original dataset to regularize the long tail, and then the recognition accuracy is significantly improved.