Inverse-Like Antagonistic Scene Text Spotting via Reading-Order Estimation and Dynamic Sampling

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Scene text spotting is challenging, particularly for inverse-like text with complex layouts (mirrored, symmetrical).
• Existing methods struggle with these irregular text orientations.

Purpose of the Study:

• To propose a unified, end-to-end trainable framework (IATS) for spotting inverse-like scene text.
• To effectively handle inverse-like texts without degrading performance on general scene texts.

Main Methods:

• Introduced an innovative reading-order estimation module (REM) using joint loss (LRE).
• Employed an initial boundary module (IBM) and boundary refinement module (BRM) for adaptive text boundary detection.
• Developed a dynamic sampling module (DSM) with thin-plate spline for improved text recognition feature sampling.

Main Results:

• Achieved superior performance on challenging scene text and inverse-like scene text datasets.
• Demonstrated effectiveness in spotting irregular and inverse-like text with high accuracy.
• The DSM proactively learned optimal features for recognition without extra supervision.

Conclusions:

• The proposed IATS framework offers a robust solution for inverse-like scene text spotting.
• The method generalizes well to various text shapes, scales, and orientations.
• This work advances the capabilities of scene text recognition systems in complex visual environments.