Classic, rule-based image processing is unbeatable for precise dimensional inspection, code reading and geometrically clearly defined features under stable conditions. Artificial intelligence (deep learning) excels with strongly variable, irregular surface defects and natural material variation. THOP AG relies on hybrid systems that combine both worlds.
When rule-based image processing dominates
Rule-based image processing works deterministically according to mathematical rules and algorithms (e.g. edge detection via Sobel operators, pattern matching via normalised cross-correlation). It requires no training data and is 100% traceable. Typical strengths:
- Precise dimensional accuracy: Determining distances, diameters, angles and tolerances down to the micrometre range.
- Identification: Robust reading and verification of 1D barcodes and 2D DataMatrix codes.
- Completeness inspection: Checking whether assembly components (e.g. screws, seals) are present at geometrically precisely defined positions.
When AI methods (deep learning) become essential
As soon as the defect scenario can no longer be mathematically defined exactly, rules reach their limits. A scratch on a brushed surface, inclusions in composite materials or irregular flash exhibit an unlimited number of visual forms. AI models (particularly convolutional neural networks, CNNs) learn these features through iterative pattern recognition from example images. Anomaly detection is particularly efficient here: the system is trained exclusively on defect-free good parts. The system understands the natural variance of the product and reliably flags any deviation from the learned good state, without needing to collect thousands of sample defective parts in advance.
Direct technology comparison
| Criterion | Rule-based image processing | AI-based inspection (deep learning) |
|---|---|---|
| Traceability | 100% mathematically verifiable (white box) | Statistical probabilities (tendency toward black box) |
| Geometric measurement | Sub-pixel accuracy in the micrometre range possible | Not suitable for precise, quantitative metrology |
| Defect variance | Requires new explicit rules for every defect pattern | Covers unpredictable, variable defect shapes |
| Data requirement at setup | Low (1–2 representative sample parts suffice) | Medium to high (requires a clean training dataset) |
| Validation effort | Simple via threshold adjustments | Requires statistical validation image sets |
Don't choose a purely “either-or” strategy. The THOPinspect software uses a hybrid approach by default: we locate the part and inspect holes or dimensions in a fully process-reliable, classic rule-based way, while complex free-form surfaces are inspected for anomalies by a lean AI model in the same inspection cycle.