Fast Recognition for Multidirectional and Multi-type License Plates with 2D Spatial Attention

Qi Liu,Song-Lu Chen,Zhen-Jia Li,Chun Yang,Feng Chen,Xu-Cheng Yin
DOI: https://doi.org/10.1007/978-3-030-86337-1_9
2021-01-01
Abstract:The multi-type license plate can be roughly classified into two categories, i.e., one-line and two-line. Many previous methods are proposed for horizontal one-line license plate recognition and consider license plate recognition as a one-dimensional sequence recognition problem. However, for multidirectional and two-line license plates, the features of adjacent characters may mix together when directly transforming a license plate image into a one-dimensional feature sequence. To solve this problem, we propose a two-dimensional spatial attention module to recognize license plates from a two-dimensional perspective. Specifically, we devise a lightweight and effective network for multidirectional and multi-type license plate recognition in the wild. The proposed network can work in parallel with a fast running speed because it does not contain any time-consuming recurrent structures. Extensive experiments on both public and private datasets verify that the proposed method outperforms state-of-the-art methods and achieves a real-time speed of 278 FPS. Our codes are available at https://github.com/qiLiu77/SALPR.
What problem does this paper attempt to address?