I chuckled at your "the worst image" sample. Which still looked quite decent all things considered.
You're "handwritten" example looks a bit "too decent" as well. I can see how that works. You first look for the edges of the table, and then you evaluate the symbol in each cell as something that matches unicode.
So, how well does this cope with increasing degradation? i.e. pencil written notes that bleed outside cell borders, curve around borders, etc.? Stamps and symbols (watermarks) across tables?