I recently did it via 'MODI' (Microsoft Office Document Imaging),
which comes by default with Office 2003, or by explicit choice in
Office 2007.
Basically its object model was to instantiate the component with a
graphics file, call OCR method of object, iterate through it's Images
collection, which would have Words collection.
Faced some issues of Interop COM wrapper around MODI component not
releasing closed files/buffers immediately.
Isolated COM wrapper in a separate process (one instance per request),
talked to this process via stdin/stdout redirected to stream, and life
was good!