CM-Net: Concentric Mask Based Arbitrary-Shaped Text Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Fast arbitrary-shaped text detection is crucial for intelligent systems but current methods face a trade-off between speed and accuracy.
• Existing real-time text detection methods exhibit significantly lower accuracy compared to non-real-time approaches.

Purpose of the Study:

• To develop a novel text detection framework, CM-Net, that simultaneously enhances both detection speed and accuracy.
• To introduce a new text representation method and a multi-perspective feature module to address limitations in current arbitrary-shaped text detection.

Main Methods:

• Proposed CM-Net framework utilizing a concentric mask (CM) representation for efficient and robust fitting of arbitrary text contours.
• Incorporated a multi-perspective feature (MPF) module to learn discriminative features from multiple viewpoints without additional computational cost.
• Introduced a multi-factor constraints loss to ensure effective learning of multi-perspective features.

Main Results:

• CM-Net demonstrated efficient and robust fitting of arbitrary-shaped text instances using the concentric mask representation.
• The multi-perspective feature module and constraints loss proved effective in recognizing discriminative text features.
• CM-Net achieved superior speed and accuracy compared to state-of-the-art real-time text detection methods on multiple benchmark datasets (MSRA-TD500, CTW1500, Total-Text, ICDAR2015).

Conclusions:

• CM-Net successfully balances detection accuracy and speed, outperforming existing real-time text detection solutions.
• The proposed concentric mask representation and multi-perspective feature learning are key to CM-Net's enhanced performance in arbitrary-shaped text detection.