Reading Arbitrary-Shaped Scene Text from Images Through Spline Regression and Rectification

Long Chen, Feng Su, Jiahao Shi, Ye Qian; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 2629-2645

Abstract


Scene text in natural images contains a wealth of valuable semantic information. To read scene text from the image, various text spotting techniques that jointly detect and recognize scene text have been proposed in recent years. In this paper, we present a novel end-to-end text spotting network SPRNet for arbitrary-shaped scene text. We propose a parametric B-spline centerline-based representation model to describe the distinctive global shape characteristics of the text, which helps to effectively deal with interferences such as local connection and tight spacing of text and other object, and a text is detected by regressing its shape parameters. Further, exploiting the text's shape cues, we employ adaptive projection transformations to rectify the feature representation of an irregular text, which improves the accuracy of the subsequent text recognition network. Our method achieves competitive text spotting performance on standard benchmarks through a simple architecture equipped with the proposed text representation and rectification mechanism, which demonstrates the effectiveness of the method in detecting and recognizing scene text with arbitrary shapes.

Related Material


[pdf] [supp] [code]
[bibtex]
@InProceedings{Chen_2022_ACCV, author = {Chen, Long and Su, Feng and Shi, Jiahao and Qian, Ye}, title = {Reading Arbitrary-Shaped Scene Text from Images Through Spline Regression and Rectification}, booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)}, month = {December}, year = {2022}, pages = {2629-2645} }