Boundary TextSpotter: Toward Arbitrary-Shaped Scene Text Spotting.

Area of Science:

• Computer Vision
• Deep Learning
• Image Analysis

Background:

• Scene text spotting, the simultaneous detection and recognition of text in images, is a challenging task.
• Existing methods often struggle with arbitrary-shaped text due to limitations of bounding box or shape mask representations.

Purpose of the Study:

• To propose an end-to-end trainable neural network, Boundary TextSpotter, for arbitrary-shaped scene text spotting.
• To develop a novel representation for text instances using boundary points that inherently encodes reading order.

Main Methods:

• Boundary TextSpotter formulates text instances as a set of ordered boundary points, enabling robust handling of arbitrary shapes.
• A single-stage text detector is integrated for efficient boundary point detection, aiming for real-time performance.

Main Results:

• The proposed Boundary TextSpotter achieves state-of-the-art or competitive results on challenging datasets (ICDAR2015, Total-Text, CTW1500).
• The method demonstrates significant improvements in inference speed compared to existing approaches.

Conclusions:

• Boundary TextSpotter effectively addresses the challenge of arbitrary-shaped scene text spotting.
• The boundary point representation offers a unified approach for both detection and recognition, enhancing efficiency and accuracy.