Detection of Oriented Scene Text in Natural Images

Abstract

Recently, in the field of computer vision, people are paying more attention to extracting and understanding text information in natural scenes, which are widely used in real life, such as image OCR, image retrieval, and multi-language translation. Text detection plays a vital role in the process of text information extraction and understanding. Diverse text patterns and highly cluttered backgrounds are the main challenges of precise text detection. This paper studies oriented scene text detection, the horizontal scene text detector, and CTPN network architecture, which is improved and extended for oriental scene text detection. The main proposed improvements include: 1. The pre-processing method and parameter setting in the post-processing are improved to make the annotation more accurate; 2. The anchor mechanism is improved by setting a series of different anchor widths depending on the anchor heights instead of the fixed width. The improved model was trained with the training set of ICDAR2015 and part of ICDAR2017 MLT. Then it was tested on the ICDAR online platform with test sets of ICDAR2013 and ICDAR2015. Finally, it was tested with some images outside the datasets. The results show that the improved model can detect text with a tilt angle more precisely.