LRANet++: Low-Rank Approximation Network for Accurate and Efficient Text Spotting.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• End-to-end text spotting unifies text detection and recognition.
• Current methods struggle with accurate and efficient arbitrary-shaped text spotting.
• A key challenge is the lack of reliable and efficient text detection.

Purpose of the Study:

• To develop an accurate and efficient end-to-end text spotter for arbitrary-shaped text.
• To address the bottleneck in text detection accuracy and efficiency.
• To propose a novel parameterized text shape representation and a triple assignment detection head.

Main Methods:

• A novel parameterized text shape representation using low-rank approximation.
• Exploiting shape correlations for a robust low-rank subspace construction.
• Minimizing an L1-norm objective for intrinsic text shape extraction from noisy annotations.
• A triple assignment detection head decoupling training complexity from inference speed.
• Integrating an enhanced detection module with a lightweight recognition branch.

Main Results:

• The proposed method enables precise reconstruction of text shapes using a few basis vectors.
• The triple assignment scheme utilizes deep sparse and dense branches for guided inference and parallel supervision.
• LRANet++ accurately and efficiently spots arbitrary-shaped text.
• Experiments on challenging benchmarks show LRANet++ outperforms state-of-the-art methods.

Conclusions:

• LRANet++ offers a superior solution for arbitrary-shaped text spotting.
• The novel detection module significantly improves accuracy and efficiency.
• The framework provides a robust and fast approach to end-to-end text spotting.