This repository contains my deep learning project where I built a Generative Adversarial Network (GAN) to generate realistic car images. The project demonstrates my journey through data preprocessing, GAN architecture development, training, and saving the generated images.
📂 Project Overview Objectives: • Build a GAN model to generate high-resolution car images. • Preprocess and merge real-world car datasets for model training. • Save and visualize generated images dynamically during training. Features: • Fully functional Generator and Discriminator architectures. • Dataset processing: Image resizing, normalization, and visualization. • Training setup with loss tracking and saving outputs for analysis.
🛠 Tools and Technologies • Python: Core programming language. • TensorFlow/Keras: Used to build and train the GAN model. • OpenCV & PIL: For image preprocessing and visualization. • Matplotlib: For plotting generated image results.
📁 Dataset The dataset consists of over 4,000 car images merged from multiple folders, resized to 128x128x3 resolution, and normalized for training. • Kaggle Dataset Link (https://www.kaggle.com/datasets/kshitij192/cars-image-dataset/data)
🧱 GAN Architecture Generator: • Converts random noise (latent vector) into high-quality 128x128x3 car images. • Includes Dense layers, LeakyReLU activations, BatchNormalization, and Conv2DTranspose layers for upsampling. Discriminator: • Distinguishes between real and fake images with a convolutional architecture. • Uses Conv2D layers, Dropout, and LeakyReLU for downsampling. Combined GAN: • The Generator and Discriminator are combined for end-to-end adversarial training. • Loss Function: Binary Cross-Entropy Loss
🚀 Training Process
- Dataset preprocessed and split into training and testing sets.
- Trained the GAN for 3000 epochs using a batch size of 64.
- Monitored training stability through Generator Loss and Discriminator Loss.
- Saved generated images dynamically during training for analysis. Results: • Successfully generated high-resolution car images within limited epochs. • Loss and accuracy metrics show promising stability for further training.
📊 Results Generated images are saved in the generated_car_images folder. Here's an example: (Add a sample image here or link a preview image.) 🌟 Future Improvements • Fine-tune model hyperparameters for better image realism. • Experiment with advanced GAN architectures like DCGAN, StyleGAN, or CycleGAN. • Train the model on a larger and more diverse dataset. 📢 Acknowledgments This project was built as part of my deep learning journey. Feel free to explore, use, and suggest improvements!
📥 Contribute Feel free to open issues or submit pull requests if you’d like to improve this project.
🔗 Links • GitHub Repository: Link • Kaggle Dataset: Link
Contact Me: www.linkedin.com/in/mohammad-junayed-ete20