[Paper Review] Image Inpainting Based on a Novel Criminisi Algorithm

Yuheng Song, Yan Hao|arXiv (Cornell University)|Aug 13, 2018

Generative Adversarial Networks and Image Synthesis8 references4 citations

TL;DR

This paper proposes an improved Criminisi algorithm for image inpainting that addresses error propagation and suboptimal patch matching. By redefining priority calculation using Manhattan distance mean and integrating Euclidean distance into matching strategy, the method enhances visual quality and reduces artifacts, demonstrating superior results in quantitative and qualitative evaluations.

ABSTRACT

In view of the problem of image inpainting error continuation and the deviation of finding best match block, an improved Criminisi algorithm is proposed. The improvement was mainly embodied in two aspects. In the repairing order aspect, we redefine the calculation formula of the priority. In order to solve the problem of error continuation caused by local confidence item updating, the mean value of Manhattan distance is used for replace the confidence item. In the matching strategy aspect, finding the best match block not only depend on the difference of the two pixels, but also consider the matching region. Therefore, Euclidean distance is introduced. Experiments confirm that the improved algorithm can overcome the insufficiencies of the original algorithm. The repairing effect has been improved, and the results have a better visual appearance.

Motivation & Objective

• To address error continuation in traditional Criminisi algorithm due to confidence update mechanisms.
• To improve patch matching accuracy by incorporating regional context beyond pixel-wise differences.
• To enhance visual quality of inpainted images through modified priority and similarity metrics.
• To reduce visual artifacts and improve structural consistency in repaired regions.
• To validate the effectiveness of the proposed method through comparative experiments.

Proposed method

• Redefine the priority calculation formula by replacing the confidence item with the mean of Manhattan distances to mitigate error propagation.
• Introduce Euclidean distance as a supplementary metric in patch matching to consider spatial context and improve matching accuracy.
• Modify the patch selection process to prioritize regions with higher structural consistency and lower uncertainty.
• Use a hybrid similarity measure combining intensity differences and spatial proximity for better patch matching.
• Update confidence values using the mean Manhattan distance of neighboring pixels to stabilize the repair process.
• Apply the improved algorithm iteratively, starting from the highest priority region to ensure coherent structure propagation.

Experimental results

Research questions

• RQ1 How can error propagation in the Criminisi algorithm be reduced during image inpainting?
• RQ2 To what extent does incorporating regional context improve patch matching quality?
• RQ3 Can a modified priority formula based on Manhattan distance improve inpainting results?
• RQ4 How does the integration of Euclidean distance affect the visual quality of repaired images?
• RQ5 Does the proposed method outperform the original Criminisi algorithm in both quantitative and qualitative metrics?

Key findings

• The proposed method significantly reduces error propagation by replacing the confidence item with the mean Manhattan distance in priority calculation.
• The integration of Euclidean distance in matching strategy leads to more accurate and context-aware patch selection.
• Visual results show improved structural coherence and reduced artifacts compared to the original Criminisi algorithm.
• The method achieves better performance in preserving textures and edges in repaired regions, as confirmed by qualitative evaluation.
• Experiments demonstrate that the improved algorithm enhances overall inpainting quality without requiring deep learning or complex optimization.
• The results confirm that the modified priority and matching strategy effectively address key limitations of the original Criminisi approach.