Representation Learning Method for Circular Seal Based on Modified MLP-Mixer

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Seal impression analysis is crucial for authentication and identification.
• Existing methods like MLP-Mixer, VGG16, and ResNet50 have limitations in capturing fine-grained details.
• A need exists for more efficient and accurate seal impression representation learning.

Purpose of the Study:

• To propose Stamp-MLP, an enhanced seal impression representation learning technique.
• To improve classification accuracy for seal surfaces, product types, and individual seals.
• To develop a model with fewer parameters and better performance compared to existing architectures.

Main Methods:

• Developed Stamp-MLP, an MLP-Mixer based technique utilizing circular seal remapping instead of patch linear mapping.
• Replaced average pooling with global attention pooling for comprehensive information extraction.
• Employed three classification tasks: seal surface, product type, and individual seal identification.

Main Results:

• Stamp-MLP achieved the highest accuracy (89.61%) in seal surface classification, outperforming MLP-Mixer, VGG16, and ResNet50 with fewer training samples.
• Achieved superior accuracy rates of 90.68% for product type and 91.96% for seal impression classification.
• Demonstrated the most efficient model with the fewest parameters (2.67 M).

Conclusions:

• Stamp-MLP offers a significant advancement in seal impression representation learning.
• The proposed method provides higher accuracy and efficiency compared to established models.
• Circular seal remapping and global attention pooling are effective strategies for seal analysis.