Back to catalogue
Project
Fake Naira Detector
CNN experiment with training and test scripts for distinguishing Naira images from synthetically altered examples.
How it’s builtHow it’s built
Fake Naira Detector · CNN that classifies Naira notes as genuine or synthetically counterfeit
I built a small Keras CNN that classifies Naira notes as genuine or counterfeit from 64x64 RGB images. I generated the counterfeit class by applying random colormaps, contrast, brightness, noise and blur to genuine notes, giving 576 images across four denominations. A Streamlit app shows training curves and the classification report and runs predictions on uploads.
- images (286 genuine, 290 counterfeit) across 4 denominations
- 576
- model parameters
- 3,937,538
- accuracy measured on the training set
- 0.92
- training epochs
- 9
Technology and what it does here
- Keras 3.3 / TensorFlow 2.16
- Conv2D(32) > MaxPool > Flatten > Dense(128) > Dense(2 softmax), Adam, categorical cross-entropy
- OpenCV + Pillow
- Preprocessor.py synthesises counterfeit notes with colormaps, noise and Gaussian blur
- Streamlit
- Multipage app with charts, sample notes and upload prediction
- scikit-learn
- Classification_report and accuracy for result/result.csv
- pandas
- Loads result CSVs and training history for the app's tables and charts
Implementation details
- Each counterfeit is derived from a real note by grayscale conversion, a random cv2 colormap, random contrast (0.5-1.2) and brightness (0.8-1.5), Gaussian noise and optional blur.
- The app loads model/fake-naira-notes-detector.h5 and shows per-class probabilities in a dialog.
Technologies
- Keras
- Image Classification
- Synthetic Data
