Indian Currency Note Recognition using CNN for the Visually Impaired

Identifying a banknote's denomination is effortless with sight, yet it is a genuine daily obstacle for people living with severe visual impairment. Indian notes differ by colour, size and motif, but under poor light, with worn or folded notes, these cues are hard to rely on — and the tactile marks on notes wear off with circulation. Existing currency-reader apps are often closed-source, unreliable on Indian notes, or need a constant internet connection. This project builds an open, on-device alternative: point the phone camera at a note, a CNN classifies the denomination, and the phone speaks it aloud. The scope is deliberately honest — six denominations, one clear task, measured on a held-out test split — which makes the dataset curation, model training and accessibility design genuinely explainable in the viva.

1. A labeled dataset is assembled: photographs of Mahatma Gandhi (New) Series notes (Rs 10–Rs 500), obverse and reverse, under varied lighting, angles, backgrounds and note conditions, supplemented by publicly available labeled note images.
2. Every image is centre-cropped, resized to 224x224 and pixel-normalised to [0, 1]; the data is split into stratified train/validation/test sets.
3. A MobileNetV2 base pre-trained on ImageNet is frozen and topped with a custom head (global average pooling, dense layer, dropout, 6-way softmax), then fine-tuned with augmentation (rotation, brightness, perspective shifts).
4. Training uses early stopping on validation loss; the best checkpoint is evaluated once on the test split to produce accuracy, per-class precision/recall/F1 and the confusion matrix.
5. The model is converted to a mobile format (TFLite) and wired into the Android demo app with the same preprocessing the training used.
6. In the app, the camera frame is classified on-device and the top denomination is spoken through text-to-speech with its confidence shown; low-confidence frames ask for a retake.

Tech stack:

• Python 3, TensorFlow · Keras
• MobileNetV2 transfer learning
• OpenCV (preprocessing & augmentation)
• NumPy · scikit-learn (metrics)
• Jupyter Notebook (training & evaluation)
• TensorFlow Lite (on-device inference)
• Android (demo app) · Web Speech API (browser demo)

Dataset & model details

• Dataset: A purpose-built labeled image dataset assembled during project development: photographs of Mahatma Gandhi (New) Series Indian banknotes (Rs 10, Rs 20, Rs 50, Rs 100, Rs 200, Rs 500), obverse and reverse, captured under varied lighting, angles, backgrounds and note conditions, supplemented by publicly available labeled Indian note images (e.g. Kaggle Indian currency-notes collections). Design target: approximately 3,000 labeled images; the exact count, class balance and 70/15/15 train/validation/test split are documented in the build report.
• Task: 6-class image classification; input = 224x224x3 RGB photograph, output = probability distribution over the six denominations.
• Model: MobileNetV2 (ImageNet pre-trained base, frozen) + GlobalAveragePooling2D + Dense(128, ReLU) + Dropout + Dense(6, softmax); converted to TensorFlow Lite for on-device inference in the Android app.
• Metrics: Test accuracy, per-class precision/recall/F1, confusion matrix. Design target: approximately 88–92% validation accuracy — a development target, not a measured claim; the report documents the actually achieved figures after the training run.

• [6-denomination CNN classifier] A MobileNetV2 transfer-learning model classifies Rs 10, Rs 20, Rs 50, Rs 100, Rs 200 and Rs 500 notes from a single photograph, with the full training pipeline reproducible from the notebook.
• [Voice announcement] The identified denomination is spoken aloud through the device text-to-speech engine ("One hundred rupees"), so the app is usable without looking at the screen.
• [Offline on-device inference] The trained model is converted to a mobile format and runs entirely on the phone — identification works with no internet connection.
• [Confidence-aware feedback] Each prediction shows a confidence score; low-confidence captures trigger a "please retake the photo" prompt instead of a wrong announcement.
• [Vibration backup cue] A short vibration pattern accompanies the voice result, giving a non-audio confirmation channel in noisy surroundings.
• [Full training notebook] Dataset loading, augmentation, model definition, training loop and evaluation in one reproducible Jupyter notebook with every hyperparameter recorded.
• [Evaluation report] Accuracy, per-class precision/recall/F1 and a confusion matrix on a held-out test split — the report documents what the model actually achieved, with no invented figures.
• [Working browser demo] A single-file web demo of the end-to-end flow (scan, analyse, voice output) is included alongside the Android app source.

• Purpose-collected labeled dataset of Indian banknote photographs (with the exact image list and split used)
• Complete training & evaluation Jupyter notebook with hyperparameter record
• Trained MobileNetV2-based CNN (Keras checkpoint + TensorFlow Lite mobile model)
• Evaluation report: accuracy, per-class precision/recall/F1, confusion matrix, failure-case analysis
• Android demo app source (camera capture, on-device inference, voice announcement, vibration cue)
• Single-file web demo of the end-to-end pipeline
• Project report PDF (background, CNN theory, dataset curation, methodology, honest results)
• PPT presentation for final review
• Viva Q&A preparation document (transfer learning, softmax, augmentation, evaluation metrics, accessibility design)