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

1. Problem Statement

The paper addresses the critical challenge of financial independence for visually impaired individuals, particularly in Bangladesh. While tactile features (raised dots) exist on Bangladeshi banknotes, they degrade over time and are unreliable for distinguishing between notes of similar dimensions. Existing solutions suffer from several limitations:

• Lack of Robustness: Most models fail in real-world scenarios involving complex backgrounds, varying lighting, occlusions, and worn or folded notes.
• Dataset Limitations: Previous datasets were often small, controlled, or lacked specific denominations (e.g., the 200 Taka note released in 2020).
• Architectural Constraints: Single-architecture models often struggle to balance high accuracy with the low computational cost required for resource-constrained devices (mobile/embedded).
• Black-Box Nature: A lack of Explainable AI (XAI) reduces user trust and hinders debugging in assistive technologies.

2. Methodology

The authors propose a novel Hybrid CNN Architecture integrated with a comprehensive evaluation framework.

A. Dataset Construction

A key contribution is the creation of five distinct datasets representing progressive environmental complexity, totaling 8,100 images across 9 denominations (2, 5, 10, 20, 50, 100, 200, 500, and 1000 Taka):

1. Custom Dataset-1: Public benchmarks with controlled white backgrounds.
2. Primary Dataset-1: Simple backgrounds captured via smartphone/webcam.
3. Custom Dataset-2: A hybrid of the above two.
4. Primary Dataset-2: Real-world scenarios with complex backgrounds, varying lighting, occlusions, and diverse angles.
5. Custom Dataset-3: The aggregated maximum-diversity dataset including augmented samples.

B. Model Architecture

The proposed system utilizes a Dual-Stream Feature Extraction pipeline followed by a Multilayer Perceptron (MLP) classifier:

1. Dual-Stream Feature Extraction:
   • Branch 1: MobileNetV3-Large (optimized for low-resource devices via Neural Architecture Search).
   • Branch 2: EfficientNetB0 (uses compound scaling for depth, width, and resolution).
   • Both networks are pre-trained on ImageNet. The classification heads are removed, and the final 20 layers are fine-tuned.
2. Feature Fusion & Processing:
   • Feature vectors from both branches (1280 dimensions each) are concatenated into a 2560-dimensional vector.
   • L2 Normalization is applied to ensure scale invariance.
   • Principal Component Analysis (PCA) reduces dimensionality to ~200–400 features while preserving 95% of the variance, lowering computational cost.
3. Classification:
   • A compact MLP with a bottleneck architecture (Dense layers with ReLU, Batch Normalization, and progressive Dropout) classifies the reduced features.
   • The final layer uses Softmax for probability output.
   • Total Parameters: ~327K (highly efficient for real-time deployment).

C. Training & Optimization Strategy

• Optimizers: A comparative study of Adam, AdamW, RMSProp, and SGD across four learning rates (0.0001 to 0.1).
• Validation: 5-Fold Cross-Validation was used to ensure statistical robustness.
• Metrics: Evaluation included Accuracy, Precision, Recall, F1-Score, Cohen's Kappa, Matthews Correlation Coefficient (MCC), and AUC-ROC.
• Explainability: LIME (Local Interpretable Model-agnostic Explanations) and SHAP (Shapley Additive exPlanations) were integrated to visualize decision boundaries and feature importance.

3. Key Contributions

1. Novel Hybrid Architecture: The first integration of MobileNetV3-Large and EfficientNetB0 for Bangladeshi currency, leveraging complementary feature extraction strengths.
2. Progressive Dataset Curation: Creation of five datasets ranging from controlled lab conditions to highly complex real-world scenarios, including the previously missing 200 Taka note.
3. Comprehensive Optimization Analysis: Systematic evaluation of four optimizers and multiple learning rates to define optimal training strategies.
4. Explainable AI Integration: Application of LIME and SHAP to provide transparency, crucial for building trust in assistive technologies.
5. Real-Time Deployment: Development of a web-based system with voice activation and auditory/textual output, demonstrating feasibility on resource-constrained devices.

4. Experimental Results

The model demonstrated superior performance across all datasets compared to baseline models (MobileNetV2/V3, EfficientNetB0, ResNet50/152, InceptionV3).

• Accuracy Performance:
  • Controlled (Custom Dataset-1): 99.75%
  • Complex Backgrounds (Primary Dataset-2): 92.84%
  • Aggregated (Custom Dataset-3): 94.98%
• Optimizer Performance: Adam with a learning rate of 0.0001 yielded the most consistent results across all datasets.
• Robustness: The model maintained high agreement metrics (Cohen's Kappa > 0.91, MCC > 0.91) even in the most challenging real-world conditions.
• Comparison: The hybrid model outperformed the best baseline (EfficientNetB0) by a significant margin, achieving a 14.15% accuracy improvement on the complex Primary Dataset-2.
• Efficiency: The model converges in 40–50 epochs with an average training time of 10 seconds per epoch on a Tesla T4 GPU, making it suitable for real-time applications.

5. Significance and Impact

• Assistive Technology: Provides a reliable, cost-effective tool for visually impaired individuals to identify currency independently, reducing the risk of fraud and exploitation.
• Real-World Applicability: Unlike previous studies limited to controlled environments, this system is validated against complex, noisy, and occluded real-world data.
• Transparency: By integrating XAI, the study addresses the "black-box" concern, allowing developers and users to understand why a specific denomination was identified.
• Scalability: The lightweight architecture (~327K parameters) ensures the system can be deployed on mobile phones and edge devices without requiring heavy cloud computing resources.

The paper concludes that the proposed hybrid approach offers a robust, accurate, and interpretable solution for currency recognition, setting a new standard for assistive financial technologies in South Asia and beyond.