SPTS-Net v2: One-Click End-to-End Scene Text ...

SPTS-Net v2: One-Click End-to-End Scene Text Spotting

Author Name : Kotaru Preethi, Madamanchi Nithin, Muvva Vignesh, Dr. M. Srikant

ABSTRACT A complete scene Text spotting has advanced significantly due to its combination with text detection and recognition. Previous approaches required manual annotations, such as horizontal rectangles, rotational rectangles, quadrangles, and polygons, which were more expensive than utilizing single-point. Our new system, SPTS v2, trains high-performing text-spotting models with a single point annotation. SPTS v2 uses an auto-regressive Transformer with an InstanceAssignmentDecoder (IAD) to predict the center points of all text instances in the same sequence. It also uses a Parallel Recognition Decoder (PRD) to recognize text in parallel, reducing the length of the sequence. The two decoders share the same characteristics and communicate using a simple yet effective information transmission technique. Extensive trials on benchmark datasets show that SPTSv2 outperforms earlier single-point text spotters with fewer parameters and 19x faster inference speed. Our SPTS v2 framework investigations indicate that single-point representation may be preferred over alternative formats for scene text spotting. This approach opens up new possibilities for scene text detecting applications that are not currently available