Chinese Text Spotter Exploiting Spatial Semantic Information in Scene Text Images
Hao Wang, Huabing Zhou · 2023
Chinese scene text spotting plays a significant role in computer vision applications as it involves extracting and interpreting Chinese information from scene images. However, existing methods often lack specialized designs for complex spatial semantic information in Chinese text images, leading to various degrees of false detections in Chinese scene text images. In this paper, we present SSTS, a novel Spatial Semantic Text Spotter for Chinese text image detection and recognition. SSTS utilizes a concise Transformer encoder and decoder architecture to effectively model spatial semantic information and predict the spatial orientation of Chinese text, reducing the likelihood of false detections caused by different combinations of Chinese characters in various orientations. Experimental results on Chinese scene public datasets demonstrate the superiority of our proposed method. In particular, SSTS achieves a new record of 78.5% in the 1-NED metric on the ReCTS dataset.