It just occurred to me...
The title, "Contents", was set in very heavy type which happened to be an unexpected edge case in the classifier and it matched the "o" with the "e" and "n" and output "Contoots".
Wouldn't it be a good idea to perform OCR - using a language model, the works - before you start classifying the JBIG2 symbols? That way, you'd have additional contextual information to say "Aha, 'contoots' is probably not what it reads here" at least in some of the cases.
Although, I realize that on "Google scale", such a complex solution could be a problem.