What is document splitting?
Document splitting is the separation of one uploaded file into the distinct documents it contains. Applicants commonly scan an entire folder in a single pass, producing a forty-page file holding a dozen certificates. Splitting establishes where each document begins and ends before classification operates on any of them.
Also called: Document segmentation
Why one upload is rarely one document
A hospital submitting an accreditation application uploads one file of eighty pages because that is what the scanner produced. Inside it are licences alongside equipment certificates and staff qualifications in no particular order. Treated as one document, the file is classified as whatever its first page resembles and extraction returns values from the wrong certificate. Splitting is unglamorous and it determines whether everything after it is working on real documents.
- A combined file is classified as whichever document appears on its first page.
- Extraction from an unsplit file returns values belonging to the wrong document.
- Applicants scan in one pass because the form permitted a single upload.
- Every later check depends on a document boundary being correct.
How document splitting works
- Page images are examined for the features that mark a document boundary.
- Candidate boundaries are proposed with a confidence value for each.
- Each resulting segment is classified, which tests whether the boundary was plausible.
- Segments that classify poorly are returned for a boundary reassessment.
- Low confidence splits are presented to a person for confirmation.
- The original file is retained unchanged with the segmentation stored beside it.
How document splitting is evaluated
Document splitting is evaluated on boundaries rather than on pages. Measure the following against a manually segmented sample:
- boundary precision and recall against manual segmentation
- proportion of segments assigned a confident class afterwards
- frequency of two documents merged into one segment
- frequency of one document divided across two segments
- share of uploads needing human confirmation of the split
Limitations, risks and common failure modes
- Repeated identical certificates are merged into one segment and counted once.
- A double-sided scan places an unrelated reverse page inside a boundary.
- Boundary detection tuned on clean scans fails on photographs of pages.
- The original file is discarded after splitting, so an error cannot be undone.