Deep Learning Or Rule-Based Vision: 3,000 Images And 10 Minutes

Overview

Industrial defect inspection has two families of method, and the argument between them is usually framed as old against new. That framing is unhelpful. Rule-based vision and deep learning solve different problems, and a production station that uses only one of them will be either unreliable on variable defects or unnecessarily expensive on simple ones.

A rule-based system measures something explicit: a threshold on grey level, an edge position, a blob area, a gradient direction. A deep learning model infers a pattern from examples. The first is predictable and explainable, the second is flexible but depends on data. The engineering question is which of those properties the defect class needs.

Big Bird Industrial (formerly Shimada Big Bird Industrial) has been building cleaning and deburring systems for automotive powertrain components since 2001, and the detection figures used here come from stations running on production lines rather than from benchmark datasets.

1. What Rule-Based Vision Does Well

Rule-based vision is a measurement, and measurement is what it should be used for. When the defect has a geometric definition, a rule expresses it exactly.

Geometric And Dimensional Defects

A hole that is missing, a boss that is 2 mm too tall, a thread that has been cross-tapped, a seal groove that is out of position: these are geometric facts. A rule-based measurement returns a number with a tolerance, and the decision is a comparison. There is no training set, no ambiguity and no drift as the part population changes.

Explainability Under Audit

A rule-based decision can be described to an auditor in one sentence, because it is a threshold on a measured value. On safety-related characteristics, and on customer audits where the inspection method has to be documented, that property is worth more than a marginal improvement in detection rate.

2. Where Rule-Based Vision Breaks

A rule works when the defect looks the same every time. It stops working when the defect varies in appearance or when the background varies around it.

Variable Appearance

Casting porosity, cold shuts, sand inclusions and flow marks change shape, size, contrast and colour from part to part. Writing a rule for each variation produces a rule set that grows until it conflicts with itself, and the conflicts appear as false calls on good parts.

Texture And Lighting Variation

Machined surfaces have tool marks, cast surfaces have grain, and both change with tool wear and with the position of the part under the light. A threshold tuned on Monday flags the whole population on Friday after a tool change, not because the process got worse but because the texture changed.

3. What Deep Learning Changes

A deep learning model does not measure a feature; it learns the distribution of appearances that belong to a class. That makes it tolerant to the variation that breaks thresholds, at the cost of needing examples and of being harder to explain.

Detection Rates In Production

On cast aluminium cylinder heads, a deep learning surface inspection station achieves a miss rate around 0.1 percent with a false-call rate of 3 to 5 percent, at a 36-second cycle and across five part variants. On transmission housings the same order of miss rate is achieved with a 5 percent false-call rate, at 58 seconds for an end cover and 112 seconds for front and rear housings. Those figures describe a system that has been tuned against a specific criterion, not a generic capability.

Supervised, Semi-Supervised And Unsupervised

Supervised training needs labelled defect examples and gives the best accuracy per class. Semi-supervised training uses a small labelled set with a large unlabelled set, which suits programmes where defects are rare. Unsupervised training learns what normal looks like and flags deviations, which is the right choice when the defect is rare or when its appearance is not yet known. A stamping line inspecting with more than twenty cameras at a 5-second cycle can run unsupervised because normal variation is well characterised and defect examples are scarce.

4. The Training Data Problem

The practical bottleneck in deep learning inspection is not the model, it is the dataset. Collecting real defect examples takes months because defects are rare, and annotating them takes weeks because the boundary of a defect is often a judgement call.

Synthetic Defects As A Shortcut

Defect generation algorithms create plausible defects on images of good parts. For a set of 3,000 images, manual defect creation and labelling takes about 20 days and 5 days respectively; the same set generated by algorithm takes about 10 minutes and 1 minute, and the synthetic labels are exact because the generation process knows where it placed the defect. The result is not a substitute for real defect validation, but it compresses the time to a working first model from months to weeks.

The validation step still needs real defects, and it needs them from the production process rather than from a sample collection, because real defects carry the surface context that synthetic ones approximate. A station is not proven until it has run against parts that failed downstream.

5. Choosing Per Defect Class

The productive approach is to assign each defect class in the inspection specification to the method that fits it, rather than choosing one method for the whole station. A station can run a rule engine for presence and position checks and a deep learning model for appearance defects on the same part, with the results fused into one decision.

That assignment also clarifies the acceptance test. Geometric classes can be verified with a gauge or a coordinate measurement; appearance classes need a defect panel with agreed examples. Writing the two verification methods separately at the specification stage avoids the situation where a supplier demonstrates a detection rate on one class and applies it to all of them.

6. Data Infrastructure And Model Maintenance

A deep learning station produces images continuously and decisions continuously. Both need to be stored, because the first question after a customer complaint is which images the station saw and what it decided. A station that keeps only the images it rejected cannot answer that question for the part that passed.

Retaining The Right Images

The practical policy is to retain all images for a short window, all rejected images for a longer one, and all images linked to a traceability identifier for the production retention period. That combination covers the audits, the root-cause investigations and the retraining set without unbounded storage growth.

Model Versions And Revalidation

Every retrain produces a new model, and a new model can change the false-call rate on classes it was not retrained for. Version control, a fixed validation set and a comparison against the previous version before deployment are the minimum discipline. Without them, an improvement on one defect class can silently degrade another, and the change is attributed to the process rather than to the model.

Frequently Asked Questions

Does Deep Learning Replace Rule-Based Vision Entirely?

No. Geometric checks remain cheaper, faster and more explainable when implemented as rules. Deep learning adds capability on appearance defects where rules become fragile, and the two are commonly combined on the same station.

How Many Defect Images Are Needed To Train A Model?

It depends on the variability of the defect class. A few hundred well-labelled examples can be enough for a consistent defect on a consistent background, while variable casting defects need several thousand. Synthetic generation is used to reach that count faster, followed by validation on real defects.

What Miss Rate Is Realistic In Production?

Around 0.1 percent on casting and machining defects, with false-call rates between 2 and 7 percent depending on the surface and the defect class. Quoting a miss rate without the corresponding false-call rate describes half of the system.

What Happens When A New Defect Appears?

The model has to be retrained with examples of the new class, which is why the platform matters as much as the model. A station that cannot add a class without a software project will accumulate false negatives silently.

Conclusion

Rule-based vision answers questions that have a geometric answer. Deep learning answers questions about appearance that no threshold can capture. Assigning each defect class to the right method is a specification decision, and making it explicitly is what separates a station that holds its performance from one that is retuned every month.

The dataset, not the algorithm, is usually the limiting factor, which is why the time to a first working model is a better measure of a supplier than the architecture they name. The second question to ask is how a new defect class is added, because every line produces a defect nobody anticipated in the first year.