🖼️ Exemplar-Based Image Inpainting

This project implements the exemplar-based inpainting algorithm as described in the paper "Region Filling and Object Removal by Exemplar-Based Image Inpainting" (Criminisi, Pérez, Toyama – IEEE Transactions on Image Processing, 2004). The goal is to remove unwanted objects or fill missing regions in an image by copying similar patches from the remaining content.

---

📌 Highlights

• 🔍 Patch-based inpainting using local similarity
• 🧠 Priority-driven filling with confidence and data terms
• 🌀 Adaptive patch search zones for better results and faster execution
• 🌄 Support for both grayscale and color natural images

---

🧪 Method Overview

The algorithm works by iteratively filling the missing region (target) using the most similar patch from the known region (source). The order of filling is controlled by a priority term defined as:

P(p) = C(p) · D(p)

Where:

• C(p) is the confidence term: favors concentric filling from known pixels.
• D(p) is the data term: encourages propagation of linear structures like edges.

The best matching patch is chosen using Sum of Squared Differences (SSD), excluding unknown pixels from the comparison.

Patch size, search zone radius, and the presence of data/confidence terms can be tuned for optimal results.

---

💡 Improvements

To improve performance:

• The patch search is limited to a window around the target region.
• Several versions were tested with and without the data term D(p) to assess its visual impact.
• Multi-resolution strategies and a prototype of PatchMatch were explored as faster alternatives.

---

🎨 Results

Below are visual examples of the inpainting process. On the left: original image with masked area. On the right: restored image.

🪂 Example 1 — Bungee Jumping Scene

Original	Inpainted

---

🌴 Example 2 — Island with Palm Trees

Original	Inpainted

---

🦁 Example 3 — Lion in the Wild

Original	Inpainted

---