2021/07/13 by Tao Sheng, Jie Chen, Sheng, Tao +3
Computer Science · Engineering · #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Handwritten Text Recognition Techniques #Image Processing and 3D Reconstruction #Image Retrieval and Classification Techniques #Vehicle License Plate Recognition
paper · pdf · doi:10.48550/arxiv.2107.05945
openalex publication_date 2021/07/13 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Scene text detection remains a grand challenge due to the variation in text\ncurvatures, orientations, and aspect ratios. One of the hardest problems in\nthis task is how to represent text instances of arbitrary shapes. Although many\nmethods have been proposed to model irregular texts in a flexible manner, most\nof them lose simplicity and robustness. Their complicated post-processings and\nthe regression under Dirac delta distribution undermine the detection\nperformance and the generalization ability. In this paper, we propose an\nefficient text instance representation named CentripetalText (CT), which\ndecomposes text instances into the combination of text kernels and centripetal\nshifts. Specifically, we utilize the centripetal shifts to implement pixel\naggregation, guiding the external text pixels to the internal text kernels. The\nrelaxation operation is integrated into the dense regression for centripetal\nshifts, allowing the correct prediction in a range instead of a specific value.\nThe convenient reconstruction of text contours and the tolerance of prediction\nerrors in our method guarantee the high detection accuracy and the fast\ninference speed, respectively. Besides, we shrink our text detector into a\nproposal generation module, namely CentripetalText Proposal Network, replacing\nSegmentation Proposal Network in Mask TextSpotter v3 and producing more\naccurate proposals. To validate the effectiveness of our method, we conduct\nexperiments on several commonly used scene text benchmarks, including both\ncurved and multi-oriented text datasets. For the task of scene text detection,\nour approach achieves superior or competitive performance compared to other\nexisting methods, e.g., F-measure of 86.3% at 40.0 FPS on Total-Text, F-measure\nof 86.1% at 34.8 FPS on MSRA-TD500, etc. For the task of end-to-end scene text\nrecognition, our method outperforms Mask TextSpotter v3 by 1.1% on Total-Text.\n