TextBoxes++: A Single-Shot Oriented Scene Text Detector.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Scene text detection is crucial for text recognition systems but faces challenges like arbitrary orientations and varied aspect ratios.
• Existing methods often require complex post-processing, impacting efficiency.

Purpose of the Study:

• To introduce TextBoxes++, an efficient and accurate end-to-end trainable scene text detector.
• To address the challenges of detecting arbitrary-oriented text in natural images.

Main Methods:

• Developed TextBoxes++, a single-pass network for scene text detection.
• Utilized an efficient non-maximum suppression for post-processing.

Main Results:

• TextBoxes++ demonstrated superior text localization accuracy and runtime performance on multiple public datasets.
• Achieved an f-measure of 0.817 at 11.6fps on ICDAR 2015 and 0.5591 at 19.8fps on COCO-Text.
• Significantly improved performance in word spotting and end-to-end text recognition tasks when combined with a text recognizer.

Conclusions:

• TextBoxes++ offers a highly accurate and efficient solution for arbitrary-oriented scene text detection.
• The method sets a new state-of-the-art for scene text recognition tasks.