Check CMR photo legibility before OCR

Freight document photos arrive from cabs, ramps and warehouses. Glare, motion blur, shadows, folds and cropping can make a CMR impossible to read. An OCR or vision model may still return a well-formed field value, which makes an unreadable source especially risky.

The freight document control case places a deterministic legibility check before extraction. The result is either a hold with a concrete capture instruction or a structured OCR request that continues to human review.

Reject unreadable inputs before extraction

Measure the image before sending it to a model. A useful check can combine:

• brightness distribution;
• contrast;
• edge or sharpness signal;
• overexposed and underexposed area;
• document coverage and framing.

Store the measurements, threshold version and reason for the hold. The driver or warehouse operator should learn what to fix: move closer, reduce glare, include the full page or retake the photo in better light.

The quality score is a routing aid. It cannot prove that every field is correct. An accepted image still needs structured extraction, source regions and an operator approval before the data enters a settlement or invoice process.

Separate image quality from field quality

The acceptance contract should state the fields required from the document: shipment identifier, parties, dates, quantities, signatures or other values used by the process. Score each field against its source region and keep an explicit “unreadable” result.

Use these routes:

The hold should not be a silent error. It should carry the measured reason, the case identifier and the action required from the person submitting the document.

Calibrate with real CMR photos

The threshold should come from a labelled set of company images. Include devices, drivers, shifts, locations, paper conditions and lighting. Ask an experienced operator which required fields can be read safely, then compare that decision with the code's route.

Review four groups:

1. accepted and readable;
2. held and unreadable;
3. accepted but unreadable;
4. held although readable.

The third group is the main data-quality risk. The fourth group creates repeat-photo requests and delays. Choose the tradeoff with the process owner, then retest after changing the camera app, threshold, document template or OCR service.

Keep the model request reviewable

Send the image with a structured extraction schema and a source-region requirement. Store the model version, request identifier, fields returned and any refusal. A reviewer should see the field beside the relevant area of the document.

Do not let an extraction write directly to the ERP or invoice record. Place it in a review queue with the source, quality result and detected conflicts. Low-confidence fields and changed identifiers should require an explicit decision.

Measure the whole document path

Model usage is only one cost. Track:

• quality holds and repeat-photo requests;
• field-level extraction corrections;
• manual review time;
• documents accepted with later corrections;
• time from capture to approved record;
• storage, provider and integration cost;
• differences by device, location and operating condition.

The business decision depends on the full case cost. A stricter check can save extraction calls while increasing contact work. A permissive check can reduce repeat captures while creating correction and dispute work.

Configure a safe handoff

Define the owner for unreadable images, conflicting fields and missing signatures. Give the person a clear task and keep the original source. Make the manual route available when the quality service, OCR provider or network is unavailable.

Run the first workflow on one document class and one downstream record. Expand after the team can show stable thresholds, source-linked fields, complete review history and an agreed response to each hold reason.

CMR readiness list

1. Which fields must be readable for the process to continue?
2. Which image signals indicate a likely unreadable source?
3. Who labelled the calibration set and approved the threshold?
4. Does each extracted field retain its source region?
5. Which conflicts require a person?
6. Can the process request a new capture with a useful reason?
7. What prevents an unreviewed field from entering the ERP?
8. What happens when OCR or the quality service is unavailable?
9. Which cost and delay measures decide whether the workflow expands?

Bring one document class to a free process scan. Syntalith can map the quality check, review queue and downstream handoff before the first OCR integration is approved.