A Stackable Attention-Guided Multi-scale CNN for Number Plate Detection

Yixuan Wang, Shangdong Zheng, Wei Xu, Yang Xu, Tianming Zhan, Peng Zheng, Zhihui Wei, Zebin Wu

Abstract: High-speed railway transportation faces various risks due to its increasing laying area. A key to find the abnormal conditions is to locate the position of a pillar nearby. The complex background and different size of number plate (NP) makes the detection of number plate in overhead catenary system (OCS) a hard work. In this paper, we propose a novel framework and solution with two main advances: (1) a stackable attention model, which can improve the robustness of the system in complex scenarios; (2) multi-scale feature fusion stage, improve detection accuracy of distance pillar number plate. Both of them can be integrated into any CNN architectures seamlessly with negligible overheads and are end-to-end trainable along with base CNNs. We demonstrate the effectiveness of our method with experiments on different high-speed train lines and one benchmark dataset – Pascal VOC [4].