End-to-End Scene Text Detection and Recognition Based on Feature Interaction

The present scene text extraction methods generally include two independent steps of text detection and text recognition, which is not beneficial to the interaction of feature information and may result in the accumulation of errors. Aiming at the above problems, an algorithm is proposed to achieve end-to-end text extraction by interacting information between text detection and recognition. In the process of extracting scene text, Swin-Transformer and CE-FPN are adopted as the base network to extract abundant feature representations at various scales which reduce the impact of information loss on text detection and recognition; Using Recognition Conversion to guide text localization by recognition loss enables feature interaction of detection and recognition; CTC is utilized in recognition to achieve the constraint of Attention decoding, which guarantees the full extraction of long text character information. Experimental results on several datasets show that the method in this paper demonstrates better performance compared to the existing typical models.