Rathnavelu/indian-currency-cnn-yolo

🇮🇳 Indian Currency Detector (CNN + YOLO) — Mobile Optimized

A lightweight, combined YOLOv8-nano + MobileNetV3-Small model for real-time Indian currency detection and denomination classification, optimized for mobile deployment.

🏗️ Architecture

Camera Frame → [YOLOv8-nano] → Detect & Localize Currency → [MobileNetV3-Small] → Classify Denomination

Component	Model	Params	ONNX Size	Purpose
🔍 Detector	YOLOv8-nano	3.01M	11.6 MB	Locate currency in frame
🏷️ Classifier	MobileNetV3-Small	1.08M	4.4 MB	Identify denomination
📱 Total	Combined Pipeline	~4.1M	~16 MB	End-to-end on mobile

📊 Performance

Metric	Value
CNN Classification Accuracy	100.00% (all 7 classes)
YOLO mAP@50	98.28%
YOLO mAP@50-95	68.95%
Inference speed (mobile est.)	~25-35ms per frame
Total model size (FP32)	~16MB
Total model size (INT8 est.)	~5-6MB

💰 Supported Indian Denominations

Currency Notes
₹10
₹20
₹50
₹100
₹200
₹500
₹2000

🚀 Usage

Quick Start — Classification Only

import torch
from torchvision import models, transforms
from PIL import Image

# Load model
model = models.mobilenet_v3_small()
model.classifier = torch.nn.Sequential(
    torch.nn.Linear(576, 256),
    torch.nn.Hardswish(),
    torch.nn.Dropout(0.3),
    torch.nn.Linear(256, 7)
)
checkpoint = torch.load("mobilenetv3_currency.pth", map_location="cpu")
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# Preprocess
transform = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])

# Predict
img = Image.open("your_currency_photo.jpg").convert("RGB")
input_tensor = transform(img).unsqueeze(0)
with torch.no_grad():
    output = model(input_tensor)
    class_names = ["₹10", "₹100", "₹20", "₹200", "₹2000", "₹50", "₹500"]
    prediction = class_names[output.argmax(1).item()]
    confidence = torch.softmax(output, 1).max().item()
    print(f"Detected: {prediction} ({confidence:.1%})")

YOLO Detection

from ultralytics import YOLO

model = YOLO("yolov8n_currency_best.pt")
results = model("currency_photo.jpg")
results[0].show()

ONNX Inference (Mobile/Edge)

import onnxruntime as ort
import numpy as np
from PIL import Image

# CNN Classifier
session = ort.InferenceSession("mobilenetv3_currency.onnx")
img = Image.open("currency.jpg").resize((224, 224))
img_array = np.array(img).astype(np.float32) / 255.0
img_array = (img_array - [0.485, 0.456, 0.406]) / [0.229, 0.224, 0.225]
input_data = img_array.transpose(2, 0, 1)[np.newaxis, ...]
outputs = session.run(None, {"input": input_data})
class_names = ["₹10", "₹100", "₹20", "₹200", "₹2000", "₹50", "₹500"]
print(f"Predicted: {class_names[np.argmax(outputs[0])]}")

📱 Mobile Deployment

Platform	Format	Command
Android	TFLite INT8	model.export(format="tflite", int8=True)
iOS	CoreML	model.export(format="coreml")
Cross-platform	ONNX	Already included!
Edge devices	OpenVINO	model.export(format="openvino")

📁 Files

File	Description	Size
mobilenetv3_currency.pth	CNN classifier (PyTorch)	4.2 MB
mobilenetv3_currency.onnx	CNN classifier (ONNX)	4.4 MB
yolov8n_currency_best.pt	YOLO detector (PyTorch)	5.9 MB
yolov8n_currency_best.onnx	YOLO detector (ONNX)	11.6 MB
config.json	Model configuration	—

🔬 Training Details

• Base Models: YOLOv8-nano (COCO pretrained) + MobileNetV3-Small (ImageNet pretrained)
• Dataset: ViratGarg/currency + ViratGarg/currency2 — 420 images, 7 balanced classes
• CNN Training: 30 epochs, AdamW (lr=1e-3), CosineAnnealing LR, label smoothing=0.1
• YOLO Training: 30 epochs, SGD (lr=0.01), mosaic augmentation, close_mosaic=10
• Augmentation: RandomResizedCrop, ColorJitter, RandomPerspective, RandomErasing, HSV jitter
• Hardware: CPU training (works on GPU too for faster training)

Training Recipe (Based on Published Research)

• YOLOBench (arxiv:2307.13901): COCO pretrain → fine-tune strategy
• HierLight-YOLO (arxiv:2509.22365): Lightweight detection architecture
• Currency CNN paper (arxiv:2509.06331): Augmentation recipe for currency images

⚠️ Limitations

• Trained on Indian currency notes only (no coins in current dataset)
• Small training set (420 images) — accuracy may vary on edge cases
• YOLO bounding boxes are pseudo-annotations (classification → detection conversion)
• For production use, recommend collecting more diverse training data with real bounding box annotations

📜 License

Apache 2.0