Image Translation and Unsupervised Domain Adaptation (UDA)

This project explores Image Translation and Unsupervised Domain Adaptation (UDA) to improve deep learning models' adaptability across datasets. It utilizes a CycleGAN for domain adaptation between the MNIST and USPS datasets, enabling effective classification without labeled target data.

---

Key Features

1. Dataset Preprocessing:

   • Resize images to 28×28.
   • Convert to greyscale.
   • Normalize pixel values to mean=0.5, std=0.5.

2. Convolutional Neural Network (CNN):

   • Baseline image classification on MNIST.
   • Achieved 68% accuracy on USPS dataset without UDA.

3. CycleGAN:

   • Translates images from MNIST (source) to USPS (target) domains.
   • Uses generators, residual blocks, and discriminators for image-to-image translation.
   • Enables training without paired datasets.

4. UDA Integration:

   • Applies CycleGAN-generated dataset for target domain training.
   • Improves USPS classification accuracy to 90% without labeled USPS data.

---

Methodology

Dataset Preprocessing

• Resize: 28×28 pixels.
• Convert: Greyscale to simplify inputs.
• Normalize: Mean and standard deviation to 0.5 for improved model performance.

CNN Model

• Two convolutional layers (ReLU activation, max pooling).
• Fully connected layer with ten outputs.
• Trained for 10 epochs on MNIST; tested on USPS.

CycleGAN

• Generator: Translates between domains using convolutions and residual blocks.
• Discriminator: Differentiates real vs. translated images using adversarial training.
• Cycle Consistency: Ensures original image reconstruction after domain translations.

UDA Workflow

1. Train CycleGAN with MNIST as source and USPS as target.
2. Generate USPS-like dataset from MNIST.
3. Train CNN on the CycleGAN-translated dataset.
4. Evaluate CNN on USPS dataset.

---

Implementation Details

Residual Block

• Addresses vanishing gradient issues.
• Two convolutional layers with ReLU and batch normalization.
• Enables learning from input-output differences.

Discriminator

• Four-layer CNN with leaky ReLU activation.
• Outputs likelihood of real vs. generated images.

Generator

• Encodes, processes through residual blocks, and decodes input.
• Uses Tanh activation to normalize pixel values.

CycleGAN Workflow

1. Forward pass: Translate source to target domain and back.
2. Cycle consistency loss ensures meaningful translations.
3. Adversarial training improves generator realism.

---

Evaluation and Constraints

Successes:

• Achieved 90% USPS classification accuracy using UDA.
• Demonstrated effective domain adaptation with CycleGAN.

Constraints:

1. Shape Consistency: Source and target datasets must share similar shapes (e.g., digits in MNIST and USPS).
2. Dataset Specificity: Each new dataset requires CycleGAN retraining, limiting scalability.
3. Potential Translation Errors: Domain translations may introduce minor inaccuracies.

---

Future Improvements

1. Automate domain alignment for datasets with greater variability (e.g., rotation, skew).
2. Enhance generalization across multiple datasets simultaneously.
3. Optimize computational efficiency of CycleGAN training.