Hierarchically Recognizing Vector Graphics and A New Chart-Based Vector Graphics Dataset.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Traditional image recognition relies on raster graphics, which face limitations like aliasing and information loss during scaling.
• Existing methods struggle with the inherent structural and hierarchical information present in vector graphics.

Purpose of the Study:

• To propose a novel object detection and classification approach leveraging vector graphics data.
• To enhance existing vector graphics recognition methods by incorporating multi-level feature learning.
• To introduce a comprehensive dataset for vector graphics detection and understanding.

Main Methods:

• Developed YOLaT (You Only Look at Text), a method processing vector graphics' textual representation using graph neural networks (GNNs).
• Introduced YOLaT++ for Multi-level Abstraction Feature Learning, analyzing primitive shapes, curves, and points.
• Created the VG-DCU dataset, featuring chart-based vector graphics, raster counterparts, and annotations.

Main Results:

• The YOLaT series demonstrated superior performance compared to both vector and raster graphics-based object detection methods.
• Achieved higher accuracy and efficiency on the challenging VG-DCU dataset.
• Validated the effectiveness of multi-level feature learning in vector graphics recognition.

Conclusions:

• Vector graphics offer significant potential for advancing image recognition tasks.
• YOLaT++ provides a robust framework for object detection and classification using vector graphics.
• The VG-DCU dataset facilitates further research in vector graphics understanding.