It doesn't work, plain and simple. Never mind what many others have said, that a greedy algorithm might work, that you can reduce it to TSP and apply heuristic

It doesn't work, plain and simple. Never mind what many others have said, that a greedy algorithm might work, that you can reduce it to TSP and apply heuristics, etc. The application of simulated annealing is also crippled: Suppose you reconstructed two slices of the image. To improve it, you have to take a column from one slice and move it to the other---but that doesn't lower the energy! What would lower the energy is to take one of the slices and attach it to the other, possibly flipping it in the process. So while it says "simulated annealing" on the lid, it's a random walk for the most part. The idea is workable, but it needs a different primitive operation: pick a range of columns, attach it in a different spot, possibly flipping it.