Key points
- Start from the defect: if it can be defined as a measurement or a rule, rule-based tools usually fit.
- Deep learning suits defects whose shape and appearance vary on a naturally variable surface.
- Neither method finds a defect the image does not show; lighting and optics come first.
- Deep learning moves effort from writing rules to collecting and labelling images, and to retraining.
- Both are validated the same way, and both can run in one VisionMaster solution.
What each approach is
Rule-based inspection is a chain of tools an engineer builds and tunes: find the part, measure an edge, count objects, read a code, compare the result with a limit. Every decision can be traced to a measured value and the limit it was compared with. When the image is stable, the result is predictable and easy to explain in an audit.
Deep learning uses a model trained on labelled images. Instead of writing rules, you show it examples of good and defective products, and it learns to classify an image, detect and locate objects, or mark defect regions pixel by pixel. Some methods learn only from good samples and flag anything unusual. The model returns a score or a region; people still set the threshold that turns it into pass or fail.
Neither method can find a defect the image does not show. If lighting and optics do not make a scratch stand out, no algorithm will catch it reliably, which is why every project starts with a sample test.
Where each approach fits
Rule-based tools fit defined, measurable faults: presence and absence, position and orientation, counting, dimensions against a tolerance, fill height on transparent containers, barcode and 2D code reading, and verification of printed text in a consistent font.
Deep learning fits variable appearance defects: scratches, marks, dents and contamination on textured or naturally variable surfaces; moulding faults that never look the same twice; difficult print for character reading; and sorting products into classes that are easier to show than to define.
Many stations combine them. Rules locate the part and read the code; a deep-learning module inspects the cosmetic area; a rule then measures the size of any defect found against the agreed limit. See surface defect detection for how this applies to scratches, marks and moulding faults.
Sample needs
Rule-based setup needs fewer images, but they must cover normal variation so limits are set where good product actually sits, plus defect samples to confirm each rule catches the fault.
Deep learning needs labelled images of each defect class across its variation. Rare defects are the hard part: if a fault appears seldom, collecting enough examples takes time. Labelling must be consistent. If two inspectors disagree about a borderline part, the model learns the disagreement. Methods trained on good samples reduce the need for defect images during training, but not during validation. The sample preparation guide explains how to build and label a representative set.
Validation
Both methods are accepted the same way. An agreed set of good and defective samples, kept separate from anything used for setup or training, is run through the station at line speed. Two figures are counted: false rejects (good products rejected) and missed defects (defects passed). The number of good and defective samples, their mix and the test speed are recorded, so each figure has a clear basis.
A small test with no misses shows the method works on those samples; it does not prove the station will miss no defects in long-term production. Borderline cases are decided with your quality team, not by the algorithm.
Changeover and retraining
With rule-based tools, a new product usually means a new recipe: teach a reference, set the limits, check on samples. Small visual changes often fit within existing limits.
With deep learning, a new product, a new material or a supplier change that alters appearance can mean collecting new images and retraining, or keeping a model per product. Plan who does this, how each model version is named and stored, and how it is re-validated before it goes live.
Maintenance over the station's life
For both methods, stable imaging is the first maintenance task: clean lenses and covers, lighting that has not aged or shifted, and mounts that have not moved. A reference set of known good and defect samples, run at agreed intervals, shows early whether results are drifting.
Rule-based stations drift when the image changes and limits no longer fit. Deep-learning stations also need an archive of labelled images, a record of model versions, and agreed triggers for retraining, such as a new product or a rise in false rejects.
Cost and effort
We do not quote figures here; effort depends on the product and the defects. The pattern is:
- Rule-based: most effort is engineering time to build, tune and document the tool chain.
- Deep learning: effort shifts to collecting and labelling images, training, validation and later retraining. Hardware may change too: HIKROBOT states that VisionMaster deep-learning inference runs on CPUs or NVIDIA GPUs, and supplies the deep-learning modules as a separate supplemental package (HIKROBOT Download Center). Check what the software scope of a quotation includes. The cost and budget guide covers the other cost drivers.
Decision table
Use these questions to find which way a check leans. The final choice is made on your samples.
| Question | Points to rule-based | Points to deep learning |
|---|---|---|
| Can the fault be described as a measurement, count or rule? | Yes: position, size, presence, a code or a fill height. | No: it is easier to show examples than to describe it. |
| How much does good product vary in appearance? | Little: consistent colour, texture and print. | A lot: natural texture, grain, moulding or print variation. |
| How varied is the defect itself? | Predictable shape and position. | Varies in shape, size, contrast and position. |
| How many labelled defect images can you collect? | Few are available. | Enough examples of each defect type, with room to collect more. |
| Must each decision be explained by a measured value? | Yes, for example against a drawing tolerance. | A score, with a highlighted region, is acceptable. |
| How often do products, materials or suppliers change? | Often, with small visual changes a limit can absorb. | Rarely, or you can plan retraining for each change. |
| Who will own the method after handover? | Technicians who adjust limits and references. | A team trained to label images, retrain and re-validate. |
Both in one VisionMaster solution
addaScan builds stations on HIKROBOT VisionMaster. We use the platform as an independent integrator; this website does not claim any official HIKROBOT partnership or distributor status. What follows is vendor-stated and describes the software, not a result on your product.
- HIKROBOT describes VisionMaster as machine vision software with algorithm tools for positioning, size measurement, defect detection and information recognition (VisionMaster product page).
- It lists deep-learning modules for classification, object detection, character positioning and recognition, and segmentation, as well as defect heat-mapping, few-sample registration learning and anomaly detection trained on normal samples only (VisionMaster product page).
- The current deep-learning supplemental package (V4.4.0) lists GPU and CPU modules for image segmentation, instance segmentation, object detection, character recognition, image classification, OCR and code reading, plus registered and edge learning. The earlier V4.3.0 package also lists anomaly detection, image retrieval and unsupervised image segmentation (HIKROBOT Download Center).
In practice this means rule-based tools and deep-learning modules can sit in one solution and one recipe, so a station does not have to choose one method for every check. The VisionMaster integration page explains what the platform provides and what we deliver around it.