Text Detection in Bilingual Scene Image Based on Improved DBNet

Scene text detection is an important task in the field of computer vision. In scene text detection, combined with the features of the text to be detected in terms of symbol shape and semantics, the detection performance of the algorithm can be effectively improved. In view of the large difference in the height of Tibetan characters, this paper proposes a method to convert rectangular labeling frames into polygonal labeling frames, so as to improve the accuracy of DBNet in the generation of probability map and threshold map. The steps of this method are as follows: first, the foreground text in the image is obtained, then the control points are sampled on the original rectangular box, the control points are moved close to the text position, and finally all the control points are connected to obtain the new polygon labeling box. In addition, in order to solve the problem of missing text pair pairing information in the existing bilingual datasets, a feature fusion method based on spatial distance and semantic distance was proposed to realize the pairing of bilingual texts in scene images. Experiments on BiTCSD showed that the proposed method can improve the accuracy of Tibetan text detection, and expand the application range of the BiTCSD after adding the pairing information.