Robust and Real-Time Bangladeshi Currency Recognition: A Dual-Stream MobileNet and EfficientNet Approach

Accessibility as a Social Justice Issue: The paper addresses a critical vulnerability for visually impaired individuals—their dependence on others to identify currency creates real risks of fraud and exploitation. This work demonstrates how AI can be designed specifically to reduce exploitation of marginalized groups rather than perpetuating it.

Robustness Requires Diverse Real-World Data: The researchers built their own dataset combining controlled scenarios with real-world conditions, then incorporated four additional public datasets. This multi-source approach achieved 92.84% accuracy on complex backgrounds versus 97.95% on controlled settings, showing that AI systems must be tested on messy, realistic conditions to be trustworthy in practice.

Efficiency Enables Equitable Deployment: By using lightweight models (MobileNetV3 and EfficientNetB0) designed for resource-constrained devices, the system can run on affordable smartphones rather than requiring expensive specialized hardware. This design choice directly impacts who can actually access and benefit from the technology.

Transparency Through Explainability: The authors incorporated LIME and SHAP—tools that explain why the AI makes specific decisions—rather than treating the model as a "black box." For an assistive technology where errors could lead to financial harm, this transparency helps users understand and verify the system's reasoning.

Rigorous Evaluation Standards Matter: The paper uses seven different performance metrics (accuracy, precision, recall, F1-score, Cohen's Kappa, MCC, AUC) and five-fold cross-validation rather than relying on a single accuracy number, providing a more complete picture of real-world reliability.