Text Detection of a Chinese Long and Dense Text Dataset

Abstract

In daily life, it is important to interpret information correctly from different types of texts. After applying the neural network to object detection, a breakthrough has been made in natural scene text detection. However, only a few research has focused on dense, long segments of Chinese text. Moreover, there are hardly any datasets that include Chinese text boxes with large gap in aspect ratio and short interval. Considering these, we propose a Network to Detect Chinese Long text (DCLnet) and a Chinese Long and Dense text Dataset (CLD). It can not only accurately detect long and dense Chinese text, but also predict text in arbitrary directions using rotated quadrilateral shapes. In this method, we have improved AdvancedEAST model. The feature extraction part selects Resnet50, one of the latest networks. Additionally, reduce-redundancy module is added before the prediction stage to reduce redundant computations. We conduct several experiments on datasets including ICDAR2015 and CLD. Through comparative analysis, the detection accuracy of the algorithm for long and dense Chinese text is obviously better than the previous methods. It achieves a competitive F-score of 0.761 on CLD dataset.