Multi-Orientation Scene Text Detection with Adaptive Clustering.

Area of Science:

• Computer Vision
• Machine Learning
• Image Analysis

Background:

• Scene text detection is crucial for image analysis but current methods struggle with non-horizontal text.
• Existing approaches often overlook the complexity of multi-orientation text in natural scenes.

Purpose of the Study:

• To develop a unified framework for adaptive hierarchical clustering to improve scene text detection.
• To propose an effective system for multi-orientation scene text detection by grouping characters adaptively.

Main Methods:

• A unified distance metric learning framework for adaptive hierarchical clustering is presented.
• The system employs sequential coarse-to-fine grouping steps: morphology-based, orientation-based, and projection-based clustering.
• Character grouping is achieved through single-link and divisive hierarchical clustering.

Main Results:

• The proposed system achieves 71 percent f-measure on the MSRA-TD500 multi-orientation dataset, surpassing state-of-the-art.
• Evaluated on multiple public datasets including ICDAR, MSRA-TD500, and NEOCR.
• A new challenging multi-orientation scene text dataset (USTB-SV1K) is introduced.

Conclusions:

• The adaptive hierarchical clustering framework effectively handles multi-orientation scene text detection.
• The proposed system demonstrates superior performance on challenging, real-world scene text datasets.
• The release of the USTB-SV1K dataset facilitates further research in this area.