I receive Searchable PDF files that comprise of scanned pages from multiple documents (structured documents such as forms and unstructured documents such as letters). As the files do not contain a separator demarcating the end of a document, I have had to resort to classifying the start page of a document (a small subset of the total pages).
Given that the class labels of interest are a small subset of the pages, I think I could improve accuracy if I could identify and classify documents rather than the start page of a document.
I am unclear on how I would go about identifying documents within the stream of pages. For now, we can assume that the document pages are sequential, and that I am able to extract text, text position and font details from the searchable PDF. I also am able to extract the images.