I'd go for the matrix-of-bools option. If you find that standard Python lists are too inefficient for this, you could use a numpy.bool array instead. Storage for a 1000x1000 pixel maze is then just 1 MB.
Don't bother with creating any tree or graph data structures. That's just a way of thinking about it, but not necessarily a good way to represent it in memory; a boolean matrix is both easier to code and more efficient.
Then use the A* algorithm to solve it. For the distance heuristic, use the Manhattan distance (distance_x + distance_y).
Represent nodes by a tuple of (row, column) coordinates. Whenever the algorithm (Wikipedia pseudocode) calls for "neighbours", it's a simple matter of looping over the four possible neighbours (mind the edges of the image!).
If you find that it's still too slow, you could try downscaling the image before you load it. Be careful not to lose any narrow paths in the process.
Maybe it's possible to do a 1:2 downscaling in Python as well, checking that you don't actually lose any possible paths. An interesting option, but it needs a bit more thought.
