Pix2Pix Image-to-Image Translation

A deep learning project using the Pix2Pix conditional GAN model to perform image-to-image translation on aerial images.

Overview

This project applies the Pix2Pix model to the TU-Graz dataset, which consists of 400 aerial images captured by a drone. The goal is to generate realistic images from label maps, optimizing performance through data augmentation and loss function modifications.

Features

• Implements the Pix2Pix GAN model.
• Uses the TU-Graz aerial image dataset.
• Baseline implementation based on existing Pix2Pix repositories.
• Enhancements:
  1. Data Augmentation (rotation, cropping, flipping).
  2. Modified Loss Function (Kullback–Leibler divergence integration).
• Performance Metrics: VIF, UQI, SSIM, and PSNR.

Performance Comparison

Below is a comparison of different configurations tested:

Model	VIF	UQI	SSIM	PSNR (dB)
Baseline	0.126	0.039	0.187	13.400
Base + KLD	0.113	0.045	0.226	13.643
Augmented	0.153	0.044	0.236	13.808
Aug + KLD	0.183	0.053	0.242	13.894

Result

For more information there is the file Report.pdf that explains the project in detail

References

• 📄 Pix2Pix Paper (Isola et al., 2018)
• 📂 TU-Graz Dataset
• 💾 Original Pix2Pix Repository

---

👥 Contributors: Stefano Iannicelli & Ettore Caputo