Image Completion Using Markov Random Field Energy Minimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image completion methods struggle with filling large holes in images, often resulting in unrealistic or artifact-filled results, especially when structures and textures are involved, and require significant computational resources and memory.
Innovation Solution
A computer-implemented method that uses a dictionary of atoms representative of the known image part, combined with Markov Random Field (MRF) energy minimization, to determine an optimal distribution of atom linear combinations for filling holes, ensuring realistic and efficient image completion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional image completion methods are used to fill large holes, then the hole can be filled, but the result becomes unrealistic and artifact-filled
Solution Approach 1:
The method performs preliminary actions by first learning a dictionary of image patches from the known regions of the image before filling the hole. This dictionary learning step prepares reusable image patterns that will be used during the completion process, ensuring realistic textures and structures are available for filling large holes without producing artifacts
Solution Approach 2:
The method copies relevant image patterns from the known regions of the image into the hole region. By learning a dictionary of patches from the surrounding areas and copying appropriate combinations of these patches into the hole, the method preserves realistic image structures and textures while avoiding the artifact-filled results of conventional methods
2Productivity
If conventional image completion methods are used, then holes can be filled, but significant computational resources and memory are required
Solution Approach 1:
The method performs preliminary dictionary learning from the known image regions, creating a compact representation of image patterns. This preliminary step allows the subsequent hole filling to use efficient patch copying and combining operations rather than computationally intensive iterative optimization, significantly reducing computational resource requirements while maintaining high productivity
Solution Approach 2:
The method changes the approach from pixel-level continuous optimization to patch-level discrete selection. By representing images as collections of reusable patches and selecting/combining these discrete units to fill holes, the method reduces computational complexity and memory requirements compared to conventional continuous optimization methods
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
The invention notably relates to a computer-implemented method for completing an image having at least one hole, the method comprising providing (S10) a dictionary of atoms representative of the known part of the image; determining (S20) an optimal distribution of atom linear combinations with a process comprising at least one minimization (S22-S28) of a respective energy defined on a respective Markov Random Field having nodes and labels, each node corresponding to a respective location of the hole and being associated to a respective label; and filling (S30) the hole with the optimal distribution. The invention improves the field of image completion.