Technology
Computer Vision
Scans, photos and layouts: what the vision layer detects before any text is read.
The concept
What the vision layer sees before anything is read
Vision works on the page as an image. It decides what can be read and how far to trust it, one region at a time.
Business documents arrive in every condition. A PDF exported from an ERP is clean. A scanned invoice is tilted and grey. A receipt photographed on a restaurant table is creased, curved and lit from one side, and a fax of a fax has lost half its contrast. Run character recognition straight on those and it returns confident nonsense. The vision layer runs first and does what a person does without noticing: turns the page the right way up, flattens it, finds where the table is, and sees that something is stamped across the total.
It returns three things. A corrected image. A map of regions with a name on each: header, line table, tax block, footer, stamp, signature, handwritten note. And a measure of image quality for each region, which becomes the starting point for the confidence on every field read from it. A faded tax line is not turned away at the door. Its fields come back with low scores and go to a person, while the clear fields on the same page carry on.
Layout is where most of the difficulty lives. A line table that runs onto a second page is one table, and the subtotal at the foot of page one is not the grand total. Two columns of text are read column by column. A stamp over a number means the number is partly hidden, and the system has to say so instead of reading around it. None of this needs a template per supplier, so a supplier you have never dealt with is read on their first invoice. Document AI covers what the reading layers do next.
Where it sits
The first of five stages, and what it passes on
PTAS AI workflows run as Perception, Intelligence, Orchestration, Action and Verification. Vision is the first stage. What it measures follows each field through the other four.
In the products
What vision handles inside DocPro, and what reaches FinSuite
The vision layer is part of every DocPro agent. FinSuite never sees the image itself, only the fields, their scores and a link to the region each came from.
DocPro
Poor scans and phone photos
Skew, curl, glare and low contrast corrected before reading. What cannot be corrected lowers the scores of the fields in that region, so a receipt is not rejected outright for one faded line.
DocPro
Tables across pages
Line tables found by their structure, rows kept as rows, and a table that continues onto the next page joined into one. Subtotals and carried-forward lines are recognised for what they are.
DocPro
Stamps, signatures and handwriting
Detected as marks before any text is read. A received stamp gives its date. A signature is recorded as present. A handwritten amendment is read where legible and flagged where it is not.
DocPro
Bundles split into documents
A scanned stack of forty receipts, or an invoice with its delivery challan behind it, is split at page boundaries before classification. An unclear boundary goes to review, and the pages stay apart until a person decides.
DocPro
Identity images
KYC images checked for glare across the photo or number, cropping and focus before fields are read. A blurred document number brings a request for a better image, and the field stays empty until one arrives.
FinSuite
The crop beside the entry
The page region each value came from travels with the record into FinSuite, so an approver looking at a posted figure can open the part of the original it was read from.
Worked example
One scanned page taken apart, then one bundle split
A tilted scan of a two-page invoice with a stamp and a handwritten correction. Then a twelve-page PDF that turned out to hold three documents, a repeat and a question. Values are illustrative.
Page 1 of a scanned invoice, before reading. Illustrative example
The page was not rejected for one bad corner. Quality was measured region by region, so the doubt stays with the fields it affects.
Straightened first
A four-degree tilt is enough to push a table's columns into each other. It was corrected before anything was read.
Quality where it matters
The faded third lowered the scores of the totals inside it. The header kept its clean start.
A mark is reported as a mark
The stamp over the total was reported as covering a value, so that value goes to a person.
Multi-page files raise a different question before any page is examined this closely: where does one document end and the next begin? A scanning room that feeds a stack through in one pass produces exactly this kind of file.
A twelve-page scan, split before anything is read. Illustrative example
Two invoices merged into one record fail later in confusing ways. Settling page boundaries first keeps each problem where it started.
Boundaries before types
Each page is assigned to a document before any document is classified.
Repeats caught early
Pages scanned twice are held as repeats before they can become a duplicate invoice.
Doubt goes to a person
Two pages that might be one receipt or two were shown to a reviewer side by side.
Limits and controls
What vision cannot recover
Vision can straighten a page and lift a faded line. It cannot restore what the scan never captured. A total under a dense stamp may stay unreadable however the image is processed, and the right output for it is a flag. Very low resolution, heavy compression and photos of a screen lose detail no correction brings back. Handwriting varies far more than print, so handwritten values start with lower scores, and we would not post a handwritten total without a person looking at it.
Vision also has no opinion on authenticity. It finds that a stamp or signature is there, where it is and what it covers. Whether it is genuine is outside what the platform decides. The controls below follow from these limits.
Quality measured per region
Image quality is scored where each field sits, so one bad corner lowers the fields in it and leaves the rest of the page alone.
Re-capture over guessing
A page or ID image below the readable floor goes back for a better copy or to a person before extraction is attempted.
Boundaries settled first
Page splits in a bundle are scored like any other decision, and an unclear one goes to review before two documents can merge.
Crop kept with the value
The region of the image each value came from is stored with it, so review and audit look at the same pixels the reader did.
FAQ
Questions evaluators ask about computer vision
Do you need a template for each supplier's layout?
No. Regions and tables are found from the structure of the page, so an invoice layout the system has never seen is read on its first arrival. What a template would have given you, the certainty that a field is where it was last time, comes instead from the scores, and those are watched by supplier. A supplier who redesigns their invoice shows up as a change in scores for that supplier, and is looked at then.
What image quality do our scanners and phone cameras need?
Better images cost less, because fewer fields end up in review, but most documents have no hard minimum. Scan at a setting where small print is easy for a person to read, avoid heavy compression, and keep the whole page in frame and in focus for phone photos. During the assessment we measure the quality of your actual intake and tell you which sources are driving review, which is usually more useful than a scanner specification.
Can it tell whether a signature or a stamp is genuine?
No. The vision layer detects that a signature or stamp is present, where it is and what it covers, and it reads a stamp's date where legible. Whether a signature is authentic is a question for a person or a dedicated verification process. The platform does not answer it, and we would not want it to appear to.
Why is a page not rejected when part of it is unreadable?
Because most of it usually is readable. Quality is measured region by region, so a faded tax block lowers the scores of the fields inside it and leaves the header and line table alone. The clear fields proceed on their own scores, and the doubtful ones go to a person with the crop of the image. Rejecting the whole page would send a person thirteen fields they did not need to see.
Keep reading
Where the vision layer shows up
Document AI
The reading layers that start from what vision hands over.
Read more →DocPro Receipt Agent
Phone photos of receipts, where image quality matters most.
Read more →DocPro KYC & Identity Agent
Identity images checked for glare and focus before any field is read.
Read more →Machine Learning
How scores, image quality included, are calibrated on your documents.
Read more →Next step
Send us your worst scan
Upload a real document, tilted, stamped or photographed on a desk, and see where the vision layer lowered the scores. Bring a sample of your intake and we will tell you which sources drive review.