Deep learning in industrial machine vision
When rule-based image processing reaches its limits.
Deep learning has long been part of everyday life: face and speech recognition, autonomous driving, medical diagnostics. In production it solves inspection tasks that cannot be described reliably with fixed rules - such as telling a tolerable cosmetic flaw from a real defect.
This page explains what deep learning actually delivers in machine vision, where classic methods remain the better choice and which components you need to get started.
What deep learning is
Deep learning is a branch of artificial intelligence. It is based on artificial neural networks that learn from many example images to recognise and classify patterns. Instead of programming inspection rules by hand, you show the system good and bad parts - it derives the distinguishing features itself.
As with human learning, experience counts: the more representative examples the network has seen, the more reliably it recognises variants that no rule anticipated.
Rule-based or learned: where the difference lies
Classic machine vision works with programmed rules: find edges, measure distances, compare grey values. It is fast, traceable and ideal for every task that can be described exactly - dimensions, positions, presence, codes.
It reaches its limits when defect patterns vary widely, surfaces and lighting fluctuate, or natural scatter has to be separated from real deviations. That is where deep learning is strong: it judges the whole image, tolerates variance and improves with every new example without reprogramming the core algorithm.
Advantages of deep learning at a glance
- Automatic feature extraction: relevant features come from the images, not from hand-written rules.
- Robust against variance: colour shifts, surface texture, slight deformation and positional offsets matter less.
- Defect classification: scratches, inclusions, contamination and shape errors can be distinguished and counted.
- OCR on difficult markings: embossed, distorted or low-contrast characters are read more reliably.
- Shorter set-up for complex inspections: collect examples instead of formulating rule sets.
- Continuous improvement: new examples extend the model, the line keeps learning.
Typical applications
- Surface and defect inspection on castings, plastics, wood, textiles and painted parts
- Completeness and assembly checks with high part variance
- Classification and sorting of natural products, food and bulk material
- OCR/OCV on packaging, labels, castings and type plates
- Presence and position checks on changing products without reprogramming
- Separating tolerable cosmetic flaws from functional defects
When classic machine vision remains the better choice
Deep learning is no substitute for good optics and lighting, and not every task needs a neural network. Rule-based methods are preferable when:
- the inspection is measurable - dimensions, tolerances, positions, distances;
- codes or plain text are read under controlled conditions;
- only a few example images of defects exist;
- every decision must be traceable in detail;
- cycle time and computing power are tight and the task runs stably without AI.
In practice many systems combine both: classic tools for position and dimension, deep learning for judging surface and appearance.
Which components you need
The easiest entry is a smart camera with deep learning tools on board: camera, lighting, evaluation and PLC interfaces in one housing, trained from example images on a PC.
Suitable devices are listed under smart cameras with AI tools - filterable by AI level, resolution and interface.
For tasks beyond a smart camera - several cameras, high resolutions, custom models - industrial cameras with PC-based evaluation are used. Image quality decides here too: a well-lit, sharp image needs a smaller model and fewer examples.
If standard software does not fit the task, we develop the evaluation - see software & AI development.
How we proceed
1. Understand the task
What has to be detected, how often, with what certainty? Which defect patterns exist, and how frequent are they?
2. Define image acquisition
Choose camera, lens and lighting so that the feature is stably visible - half the battle for any model.
3. Collect examples and train
Capture good and bad parts, train the model, validate on held-back images.
4. Integrate and secure
Connect to the PLC or control system, set thresholds, release - and plan how new variants enter the model.
The data decides
A deep learning model is only as good as its examples. Collecting images of good and bad parts under production conditions before the project starts saves weeks later.
We help with image acquisition, assess feasibility on your parts and say openly when a classic method solves the task more simply.
An inspection task for deep learning?
Send us good parts, bad parts or example images - we will narrow down whether and how the task can be solved.
Request a feasibility study