Building Hierarchical Representations for Oracle Character and Sketch Recognition

Area of Science:

• Computer Vision
• Artificial Intelligence
• Pattern Recognition

Background:

• Oracle characters are ancient Chinese hieroglyphs crucial for historical linguistic studies.
• Current sketch and character recognition methods face challenges with complex visual representations.
• Existing models often struggle to capture both low-level textures and mid-level structures effectively.

Purpose of the Study:

• To develop a robust and effective representation for oracle character and general sketch recognition.
• To address limitations in current visual representation techniques for image analysis.
• To enhance the performance of machine learning models in recognizing complex visual data.

Main Methods:

• Collected and analyzed a dataset of oracle characters.
• Evaluated existing visual representations and identified design criteria.
• Proposed a novel hierarchical representation combining Gabor-related low-level and sparse-encoder-related mid-level features.
• Integrated the proposed representation with Convolutional Neural Network (CNN) models.

Main Results:

• The proposed hierarchical representation significantly improves oracle character recognition.
• The representation demonstrates high effectiveness in general sketch recognition tasks.
• Combining the novel representation with CNNs yields superior performance compared to individual approaches.
• The hybrid model achieved superhuman performance in general sketch recognition.

Conclusions:

• The novel hierarchical representation is effective for both historical character and general sketch recognition.
• This approach offers a complementary advantage when integrated with deep learning models like CNNs.
• The developed method sets a new benchmark, surpassing human capabilities in sketch recognition.