👁️ Synthetic Iris Image Generation using GANs This project explores the use of Generative Adversarial Networks (GANs) to generate high-resolution synthetic iris images, with a focus on evaluating their potential to spoof biometric authentication systems.
📁 Project Overview The notebook performs the following tasks:
Extracts and processes a dataset of iris images.
Defines a data generator for grayscale image batching and preprocessing.
Trains a GAN (using a custom train_gan function assumed to be defined elsewhere).
Evaluates both the discriminator and generator models using classification metrics.
Visualizes generated images.
Packages and downloads the output results.
🛠️ Technologies & Libraries Used Python 3 (Colab)
NumPy
OpenCV (cv2)
Keras (assumed usage in GAN training)
Matplotlib
scikit-learn
Google Colab APIs (for file handling and download)
📂 Dataset The dataset is stored in a ZIP file (dataa.zip) and contains grayscale iris images in multiple formats (.png, .jpg, etc.). The structure is recursively read to extract all valid image files.
🚀 How it Works
-
Extract Dataset python Copy Edit zipfile.ZipFile.extractall() Unzips the dataset for further processing.
-
Data Generator Class DataGenerator inherits from keras.utils.Sequence to:
Batch grayscale iris images
Resize them to 64x64
Normalize pixel values to [-1, 1]
-
GAN Training (Assumed Pre-defined Function) python Copy Edit train_gan(generator, discriminator, gan, data_generator, epochs=5000, batch_size=32, save_interval=500) Trains the generator and discriminator on batches from the dataset.
-
Evaluation Discriminator: Evaluated on real + fake images using classification_report.
Generator: Fake images are classified by the discriminator; performance is again evaluated using a confusion matrix-style report.
-
Visualization The best generated images are displayed using matplotlib.
-
Output Generated images are saved, zipped, and automatically downloaded via:
python Copy Edit !zip -r generated_images.zip generated_images/ files.download("generated_images.zip") 📊 Evaluation Metrics Accuracy, precision, recall, and F1-score using sklearn.metrics.classification_report
Separate reports are generated for the discriminator and generator.
🖼️ Sample Output Visual outputs from the generator are grayscale iris images generated from random noise vectors.
📌 Note This code assumes that the following models and methods are defined elsewhere:
generator: Keras/TensorFlow-based image generator
discriminator: Binary classifier for real vs fake images
gan: Combined model for training the generator
train_gan(...): Custom training loop for adversarial training
📥 Setup Instructions Upload your dataset as a ZIP file (dataa.zip) to your Colab session.
Ensure the generator, discriminator, and GAN models are defined.
Run all cells to preprocess data, train the GAN, and evaluate results.