Overhead catenary system (OCS) images exhibit great variations with clutter backgrounds, complex scenes and oblique views which pose great difficulty for automatic pillar number plate (NP) detection and recognition (NPDAR). Although these tasks have an important and practical significance for railway transportation, little researches have been done on these fields. In this paper, we propose a complete automatic NPDAR system with two main advantages: (1) For detection task, we propose Skip Connection Attention Module (SCAM) for adaptive feature refinement. Based on SCAM, the Attention_Guided Feature Fusion (AFF) module is designed for building high-level feature maps at different scales. (2) A novel convolution module, width/height convolution module (W/H-CM) was designed for NP recognition to capture global feature information efficiently. The W/H-CM extracts contextual information from two other perspectives compared to common convolution operation and iteratively generates supplementary information, making the representation of features more comprehensive. Both of them can incorporate with state-of-art feed forward network architecture in an end-to-end training fashion. We conduct extensive experiments on both our datasets and standard benchmarks PASCAL-VOC, MS COCO to verify competitive performance of our method.

中文翻译：

---

无约束场景下的支柱车牌检测与识别

悬链线系统（OCS）图像呈现出巨大的变化，背景杂乱，场景复杂，视角倾斜，这给自动支柱号牌（NP）检测和识别（NPDAR）带来了很大的困难。尽管这些任务对铁路运输具有重要的现实意义，但在这些领域的研究却很少。在本文中，我们提出了一个完整的自动 NPDAR 系统，具有两个主要优点：（1）对于检测任务，我们提出了用于自适应特征细化的跳过连接注意模块（SCAM）。Attention_Guided Feature Fusion (AFF) 模块基于 SCAM，旨在构建不同尺度的高级特征图。(2) 为NP识别设计了一种新颖的卷积模块，宽度/高度卷积模块(W/H-CM)，以有效地捕获全局特征信息。W/H-CM与普通卷积操作相比，从另外两个角度提取上下文信息，并迭代生成补充信息，使特征的表示更加全面。它们都可以以端到端的训练方式与最先进的前馈网络架构相结合。我们对我们的数据集和标准基准 PASCAL-VOC、MS COCO 进行了广泛的实验，以验证我们方法的竞争性能。