The rapid development of scene text detection shows us the need for text recognition in a scene image. Road signs recognition, reading the scene image for machine translation, text recognition on commercial products, billboards, and vehicle plates are examples of text recognition on natural images. This review discussed the related research of scene text detection in the last ten years. We analysed the past strategies in scene text detection, the strengths and the weaknesses of each method. Additionally, we showed the relationship between text detection methods before and after using deep learning. Scene text detection has evolved to detect horizontal text, multi-orientation, multilingual, curved text, and arbitrary-shaped text. Researchers have proposed various methods to address this need. We evaluate the capability of the proposed framework based on the testing results of several representative benchmark datasets. This review aims to obtain opportunities or proposals to improve the existing accuracy, speed, or generalization cases (the various condition of the text appearances). We present future trends for scene text detection research to complete the review.

场景文本检测的快速发展向我们展示了对场景图像中文本识别的需求。道路标志识别、机器翻译读取场景图像、商业产品、广告牌和车牌上的文本识别都是自然图像文本识别的例子。本文综述了近十年来场景文本检测的相关研究。我们分析了过去场景文本检测的策略，每种方法的优缺点。此外，我们展示了使用深度学习前后文本检测方法之间的关系。场景文本检测已经发展到检测水平文本、多方向、多语言、弯曲文本和任意形状的文本。研究人员提出了各种方法来满足这一需求。我们根据几个有代表性的基准数据集的测试结果来评估所提出框架的能力。本次审查旨在获得机会或建议，以提高现有的准确性、速度或泛化案例（文本出现的各种条件）。我们提出了场景文本检测研究的未来趋势以完成评论。