Reproduce “Adinkra Symbol Recognition using Classical Machine Learning and Deep Learning” on OpenResearch

About this paper

Michael Adjeisah, Kwame Omono Asamoah, Martha Asamoah Yeboah, Raji Rafiu King, Godwin Ferguson Achaab, Kingsley Adjei

This research from Bournemouth University, Zhejiang Normal University, and Huizhou University introduces a comprehensive approach for automatically identifying and classifying Adinkra symbols, a visual language from Ghana. The study compiles a substantial dataset of 174,338 images and demonstrates that a fine-tuned ResNet model achieved 97.94% accuracy in symbol recognition.

Key insights

• Deep learning models, particularly transfer learning with pre-trained architectures, demonstrated superior performance compared to classical machine learning approaches for Adinkra symbol recognition.
• The created dataset of 174,338 Adinkra symbol images represents the largest known digital repository for machine learning applications in this domain.
• Model interpretability analysis using heatmaps revealed that deep learning models effectively focus on the distinctive curves and edges of the symbols, aligning with visually important features for classification.

Summary by alphaXiv.