Scene Uyghur Text Detection Based on Fine-Grained Feature Representation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Pattern Recognition

Background:

• Scene text detection is challenging due to complex backgrounds and multi-scale text.
• Existing methods struggle with similar textures, background noise, and detecting text in natural environments.

Purpose of the Study:

• To propose a multi-directional scene Uyghur text detection model.
• To enhance feature extraction and fusion for improved multi-scale text representation.
• To suppress false positives from text-like objects in natural scenes.

Main Methods:

• Utilized hierarchical residual convolutional groups for feature extraction, capturing fine-grained details and increasing receptive fields.
• Implemented an adaptive multi-level feature map fusion strategy to address information inconsistencies.
• Developed a model specifically for Uyghur text detection in complex natural scenes.

Main Results:

• Achieved a 93.94% F-measure on a self-built Uyghur dataset.
• Attained an 84.92% F-measure on the ICDAR2015 dataset.
• Demonstrated improved accuracy and reduced false positives in Uyghur text detection.

Conclusions:

• The proposed model effectively handles multi-scale and long-glued text detection in natural scenes.
• The fine-grained feature representation and spatial feature fusion enhance detection performance.
• This approach significantly improves Uyghur text detection accuracy and robustness.