Bring grayscale images to life with deep learning!
Our project transforms black-and-white images into vibrant, colorized versions using a Conditional Generative Adversarial Network (cGAN). We combine a ResNet18 encoder, Dynamic UNet decoder, and PatchGAN discriminator to learn realistic mappings from grayscale to color. Our work automates a process that once required manual skill — making it instant, scalable, and accessible.
Grayscale images are emotionally powerful, but lack visual richness. Manual recoloring is slow, costly, and inconsistent.
We train a deep learning model to learn color patterns from real-world images and apply them to black-and-white inputs — offering fast, accurate, and beautiful colorization.
- Revive old family photos
- Recolor historic archives
- Help artists and content creators
- Restore vintage films
- Create colorful educational materials
- Encoder: ResNet18 (pretrained)
- Decoder: Dynamic UNet
- Discriminator: 70x70 PatchGAN
- Loss: L1 Loss + GAN Loss
- Image Size: 256x256
| Setting | Details |
|---|---|
| Dataset | Subset of ImageNet (~10k images) |
| Libraries | PyTorch, FastAI, OpenCV, NumPy |
| Training Time | ~10 hours |
| Optimizer | Adam (β₁ = 0.5, β₂ = 0.999) |
| Evaluation Metrics | MSE, SSIM, Visual Comparison |
Top 10 good predictions based on Peak Signal-to-Noise Ratio Metric
Top 10 worse predictions based on Peak Signal-to-Noise Ratio Metric
- Real-time demo
- Mobile + web support
- Adjustable color warmth/vibrancy
- Add Vision Transformers + Diffusion models
- Tune for specific photo types (portraits, landscapes)
#UMassBoston #ImageColorization #DeepLearning #CGAN #Rami #Avanith #Pranathi