Zoom Text Detector.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Recent text detection methods prioritize speed over accuracy.
• Shrink-mask-based strategies create a dependency on mask quality, leading to performance issues.
• Existing methods suffer from feature defocusing, detail loss, and false positives, degrading shrink-mask reliability.

Purpose of the Study:

• To develop a novel text detector that achieves comprehensive performance (speed and accuracy).
• To address the limitations of current shrink-mask-based text detection approaches.
• To introduce a Zoom Text Detector (ZTD) inspired by camera zoom principles.

Main Methods:

• Introduced a zoomed-out view module (ZOM) for coarse-grained optimization in early layers, mitigating feature defocusing.
• Incorporated a zoomed-in view module (ZIM) to enhance margin recognition and prevent detail loss.
• Designed a sequential-visual discriminator (SVD) to suppress false-positive samples using sequential and visual features.

Main Results:

• The proposed ZTD demonstrates superior comprehensive performance compared to existing methods.
• Experimental results validate the effectiveness of ZOM, ZIM, and SVD in addressing identified limitations.
• The ZTD achieves a better balance between detection speed and accuracy.

Conclusions:

• The ZTD effectively overcomes the drawbacks of traditional shrink-mask-based text detectors.
• The novel modules (ZOM, ZIM, SVD) contribute significantly to the improved performance of the ZTD.
• The ZTD represents a significant advancement in text detection technology.