Scene Text Detection and Recognition Based on Iterative Correction
Li Wei Xiong, Ziyan Gui, Ying Ou, Wenxia Xu · 2022
There are a number of challenges for text detection and recognition because of the complexity of text presentation in natural scene images. This paper proposes a scene text detection and recognition model based on iterative correction to reduce the impact of the curvature of complex text on text detection and recognition. The model comprises a rectification network and a recognition network. An innovative rectification network is developed which employs a novel Bezier curve to predict the bounding rectangle of the scene text, and uses TPS transform to iteratively refine the correction of image. The recognition network adopts ResNet added to Bi-LSTM for encoding and LSTM with attention mechanism for decoding. Experiments on several benchmark datasets demonstrate that the proposed method can effectively improve the text recognition rate, especially on complex curved text.