Scene text detection plays an important role in many real-world applications. In this paper, we propose a multi-oriented scene text detection framework, which includes three main modules. We utilize a deep residual network in the front of the framework to learn text representations. A set of fixed-width, multi-ratio rotation anchors is introduced to slide over convolutional feature maps and generate the text proposals with orientation information. An in-network recurrent architecture is then seamlessly connected, where the sequential context of proposals is encoded in order to facilitate the construction of text lines. Extensive experiments are conducted on two ICDAR benchmarks to demonstrate the effectiveness of our approach.

场景文本检测在许多实际应用中发挥着重要作用。在本文中，我们提出了一个多面向场景文本检测框架，它包括三个主要模块。我们利用框架前面的深度残差网络来学习文本表示。引入了一组固定宽度、多比率旋转锚点来滑过卷积特征图并生成带有方向信息的文本建议。然后无缝连接网络内循环架构，其中对提案的顺序上下文进行编码，以促进文本行的构建。在两个 ICDAR 基准上进行了大量实验，以证明我们方法的有效性。