Skip to content

17/02/2026 4:21 PM - PACES

⬅️ [16/02/2026 12:06 PM - PS2](<./16_02_2026 12_06 PM - PS2.md>) | ⬆️ [EECS 504](<./README.md>) | [18/02/2026 4:30 PM - Lecture](<./18_02_2026 4_30 PM - Lecture.md>) ➡️

Aidan Dempster - adempst - 6125 9596

Problem
Predicting what the original image was given an image corrupted by noise.

Approach
Using an iterative approach to perform MAP inference where the maximum of the posterior should be the restored image. They use the equivalence between the Gibbs distribution and an MRF to define an energy function that will be minimized when at the maximum of the MAP. They then apply simulated annealing to actually find the optimum.

Contribution
This is a novel application of simulated annealing. They propose an MRF that models edges such that the MAP inference will place edges at the places where humans would see an edge.

Evaluation
They prove that simulated annealing will converge to the MAP given a specific temperature schedule. Although they do not actually use this schedule during inference.
They further provide examples of the algorithm working on synthetic images with known noise characterisitics, hand-drawn images, and an image of a road sign that is very blurry.

Substantiation
Although there is no direct comparison to other methods, the evaluation they do provide is relatively convincing. They show the algorithm acting on multiple types of images with different noise characterisitics and in each case the images appear substantially restored.

07 Images as Graphs PR.pdf


⬅️ [16/02/2026 12:06 PM - PS2](<./16_02_2026 12_06 PM - PS2.md>) | ⬆️ [EECS 504](<./README.md>) | [18/02/2026 4:30 PM - Lecture](<./18_02_2026 4_30 PM - Lecture.md>) ➡️