Photo defect detection for products

Defects flagged before the item ships to a customer

An inspector photographs the item with a phone or places it under a camera at the QC table. The model outlines the defects, names their type and suggests a grade by your rules. When it is unsure, the photo goes to a person, and that decision trains the model further.

How defects slip through inspection

  • By the end of the shift the inspector is less attentive. A small chip on a corner passes and comes back as a complaint.
  • Two inspectors give the same item different grades. Arguing is pointless: the standard exists only in their heads.
  • The senior inspector is on leave, and acceptance of a batch waits until they are back.
  • A customer sends a photo of a defect, and you have no photo of the item before shipping. No way to prove it was damaged in transit.

What changes after launch

  • 01Every item is checked against one standard, not by eye
  • 02Grade and defects are recorded on a photo, not by word of mouth
  • 03You see which line and which shift produce more defects

How the system works

  1. 01

    Agree on what counts as a defect

    Together with your QC team we list the defects and the grading rules: which chip is acceptable, which stain moves an item to grade 2. Without this there is nothing to train the model on.

  2. 02

    Set up the shot

    A phone in a holder, a tablet on a stand or a camera above the inspection table. We fix the lighting, background and angle, because they matter most for how reliably the model sees a defect.

  3. 03

    Train on your products

    We label your photos of good and defective items and set thresholds with your process engineer. If a rare defect has too few examples, we say so plainly and collect them during the pilot.

  4. 04

    Check every photo

    The employee takes a photo and immediately sees boxes around defects, their type and a suggested grade. A photo with no issues moves on with no extra steps.

  5. 05

    Pass doubts to a person

    When the model is unsure, the photo goes to an inspector with a flag. Their decision is saved and used for retraining, so doubtful cases become rarer over time.

  6. 06

    Keep a log

    Every photo is stored with its markup, grade, shift and batch. Managers see which defects show up most often and where.

Which scale is yours?

One task, three different projects. Pick the one that looks like you and press “I want this” — your request arrives tagged, and we send an estimate in 1–2 days.

Small business

common start

One workshop or line, one product type, and quality rests on one or two experienced inspectors.

What we build

  • One product type and an agreed list of defects
  • Capture in a phone app or a simple web form
  • Model trained on your photos, thresholds set with your QC team
  • Inspection log with photos and Excel export
  • Review of doubtful cases in the first month
Integrations
Telegram alerts to the shift lead, Excel export
Where it runs
Cloud: no infrastructure needed on your side
Timeline
Guideline: 4–8 weeks to the first working version

Mid-size business

common start

Several product lines or inspection points, each with its own defects, and defect data needs to reach your ERP (1C) or MES.

What we build

  • Models for several product types
  • Fixed capture stations on each line
  • Grade and defects written to the batch record in 1C or MES
  • Reports by defect type, shift and line
  • Support contract: we retrain the model for new products or packaging
Integrations
ERP (1C), MES, Telegram, corporate email
Where it runs
Cloud or your own server
Timeline
Guideline: 2–4 months to the first working version

Enterprise

Several plants, one quality standard, and a security rule that product photos never leave the perimeter.

What we build

  • One catalogue of defects and grades for all sites
  • Roles: operator, inspector, process engineer, manager
  • Audit log: who changed a grade and why
  • Models on your servers, updated on a set schedule
  • Service under an SLA
Integrations
1C:ERP, MES, internal quality systems via API
Where it runs
Your servers or a closed perimeter: photos and models stay inside
Timeline
Guideline: 4+ months, starting with a pilot at one site

Pilot from RUB 500,000. Beyond that we price by task and hours.

Questions

How many photos do you need to train the model?

It depends on the product and the defects. Your QC team usually collects common defects within a few weeks of normal work; rare ones have to be gathered separately. During the assessment we look at your photos and tell you what is missing.

Will the model replace the inspector?

No. It takes over the routine: clearly good and clearly defective items. Doubtful cases stay with a person, and their decisions help the model learn.

Lighting in our workshop changes all the time. Is that a problem?

Yes, lighting and angle affect the result more than anything else. That is why we start with a capture station: a lamp, a plain background, a phone holder. It is simpler and cheaper than training a model for any conditions.

What if a new defect or product appears?

The model will not recognise it on its own. We collect examples, label them and retrain. For mid-size and enterprise projects this is part of the support contract.

A quote in 1–2 days

Name the process that eats time. You get a pilot range, not a 40-slide deck.

Or message us