Classical computer vision is usually the better business choice in four situations: the task can be written down as rules (measure this gap, find this edge, check this colour), the scene is controlled, the system has to run on a CPU or a small embedded board, or someone must be able to explain every pass and every fail. Deep learning is the better fit where lighting, backgrounds or camera angles change, or the pass criterion is something a person judges by eye instead of measuring. Many production systems use both, and each does the part it is good at.
Classical computer vision versus learned models: two ways of encoding what matters
In classical computer vision, the engineer defines the features that matter. It covers techniques that do not rely on machine learning: engineered feature descriptors, algorithms with a theoretical basis, search methods and proven mathematical properties. In practice that means colour thresholds, edge filters, geometric transforms, template matching and morphological filters, all encoding domain knowledge as handcrafted rules. Classical keypoint algorithms such as SIFT, SURF, ORB, FAST, BRISK and AKAZE belong here too. So does the histogram of oriented gradients (HOG). If you want the mechanics of how those descriptors are built, our piece on feature extraction for computer vision covers them.
Deep learning flips this round. A convolutional network or vision transformer learns its features from example images, so its performance depends heavily on the quality of the labelled ground truth it is trained on. Keypoint models such as SuperPoint detect and describe features end to end, in a single forward pass.
On the factory floor the classical side often goes by the name machine vision: a rule-based inspection stack built from industrial cameras and deterministic filters running on a PLC or an embedded PC. The labels matter less than the underlying question. Can you write down the rule, or do you need a model to learn it?
When should a business use classical computer vision instead of a deep-learning model?
Use classical computer vision when the task rests on clear visual rules and the imaging conditions stay stable. The table below sets out five questions to check a project against.
| Question | Points to classical computer vision | Points to deep learning |
|---|---|---|
| Is the environment controlled? | Fixed lighting, fixed background, predictable object appearance | Conditions change across sites, times of day, camera angles or operating conditions |
| Can the task be stated as visual rules? | It depends on colour, edges, contours, size, shape or position | It depends on texture, context or semantic understanding |
| What compute is available? | CPU-only or tightly resource-constrained hardware | A GPU, NPU, edge accelerator or cloud capacity is available |
| How much does interpretability matter? | Every step must be inspected, explained and debugged | Strong performance across messy conditions matters more than interpretability |
| Would a hybrid work? | Classical handles preprocessing, region selection, measurement or filtering | Deep learning handles detection, classification or segmentation |
Two machine-vision vendors draw the same line. MVTec positions classical methods for precise, well-defined work such as measurement, positioning and barcode reading, and brings in deep learning once objects become variable or hard to define. Cognex describes rule-based tools as effective for location, measurement and orientation on very consistent targets, and deep learning as strongest on complex defects and highly variable images.
When the answers are mixed, start with the simplest method that can meet the requirement and move up only when the task requires it.
What the benchmarks show about classical speed and accuracy
Recent evidence comes from a benchmark of thirteen mirror-symmetry scoring methods (eleven hand-crafted, two reading out learned features). In that benchmark a tuned HOG descriptor scored a crop in 0.5 ms, more than 300 times faster than the best deep method at about 170 ms per crop, while trailing it by only 0.03 skill. On the single-axis leaderboard the deep readout led at 0.83 mean skill, with a CNN-filter measure (AlexNet-C2) at 0.81 and HOG at 0.80, the last two not statistically separable. AlexNet-C2 needed 10 ms per crop, about 20 times HOG’s cost, for no statistically separable gain.
The authors read this as making HOG the default choice, with a deep readout justified only when the final +0.03 is worth two orders of magnitude more compute. They are also explicit about the scope: none of the methods was trained for the task, so whether a task-trained deep head could widen the gap was not tested. On symmetry scoring, against frozen deep features, a tuned classical descriptor gave up 0.03 skill for a roughly 300-fold saving in CPU cost.
An earlier keypoint study points the same way. It compared 43 classical detector-descriptor combinations with two deep models (LF-Net and SuperPoint) on verification, matching and retrieval under viewpoint and illumination changes. Some classical combinations outperformed the pretrained deep models, and the authors found that deep models did not beat classical ones by a large margin. SIFT descriptors appeared in all three of the most successful combinations. SuperPoint was the fastest method, followed by ORB, but SuperPoint ran on a GPU, a comparison the authors themselves call unfair.
The business case: compute, data, validation and edge hardware
Hardware and running cost
Classical methods run comfortably on a standard CPU, while the deep approach in a 2021 comparative study could only reach real time on a GPU, which both raises platform cost and rules out some embedded systems. Traditional pipelines often run on CPUs or lightweight edge hardware, and even when a deep model’s inference can run on an optimised edge device, its training usually needs a GPU.
The cost argument here rests on the compute and data evidence above. In our judgement, the line items that tend to decide it are the inference hardware per camera, the effort to collect and label training data, and the cost of revalidating whenever the product or the process changes.
No training set to build
A classical approach does not require training resources. The deep method in the same 2021 study depended on a suitable multispectral database with matching aligned images, and its accuracy depended on how much information that dataset contained. That matters most where defects are rare: a 2025 analysis of medical device inspection notes that natural data scarcity can undermine the statistical robustness the EU AI Act expects of training and validation data.
Robustness when the imaging changes
On changes of imaging domain, the 2021 study’s evidence favours the classical approach. In that study, a deep descriptor trained only on visible-to-near-infrared pairs failed completely on another spectral pairing, while the handcrafted descriptor kept delivering good visible-to-mid-wave-infrared alignment. A classical edge detector designed around simple assumptions also stayed accurate on noisy natural images outside its design domain. The trade-off cuts both ways. The deep approach was better on images resembling its training set and handled geometric variation well, while the classical approach coped better with different imaging modalities.
Explainability and validation in regulated inspection
Classical algorithms rest on theory that can quantify the limits of what they can do, which the studied deep solutions could not offer, and the 2021 authors flag that lack of transparency as a barrier in sensitive applications. That matters for validation in medical device inspection. Under MDR and FDA frameworks, validating a rule-based function (for example, one that classifies on the average brightness of a defined pixel region) needs only a small sample of boundary cases. A deep model works on latent features, so assuring it needs training and validation datasets that are statistically significant and representative of future inspections.
We prefer modular, transparent and observable vision systems as a default, because when an inspection starts rejecting good parts you need to know which stage failed and why.
Measurement, alignment and gauging: the tasks classical tools were built for
Classical tools are ideal for precise, well-defined measurement work. MVTec’s HALCON 1D measuring tool measures distances and angles along lines and arcs. It projects equidistant lines perpendicular to a profile line, averages the grey values along each, and locates edges in the resulting profile. The engineer controls the trade-offs directly: a wider region of interest gives a less noisy profile when edges run perpendicular to it, and bicubic interpolation is the most accurate option but also the slowest. HALCON’s fuzzy measure object goes further, adding weighting functions to edge selection so that a reflection on a pin does not corrupt a width-and-spacing measurement.
Cognex’s Caliper tool does a similar job, measuring object widths, edge and feature locations, and the spacing of edge pairs. Simpler checks fit too: HSV thresholding can reach very high accuracy for cap detection on bottles under controlled lighting, and direct colour matching can verify label print colour or colour coverage without any model training. Vision inspection more broadly spans tasks from missing solder joints on PCBs to pharmaceutical packaging verification, with classical edge detection at one end of the toolset and detectors and segmentation networks at the other.
These tasks share a pattern. The answer is a number or a geometric fact.
Where does classical computer vision fall short?
Classical computer vision loses accuracy when conditions depart from what its rules assume. The failure modes are specific enough to check for before you commit:
- Geometry the tool was not designed for. HALCON’s measure objects assume straight edges roughly perpendicular to the profile line, and give wrong results on curved edges. Cognex’s Caliper needs the approximate location and characteristics of the edge in advance and is not meant for working out object shape.
- Lighting drift. Some HALCON edge-selection modes give a measured position that depends heavily on image brightness, so it shifts when the illumination changes.
- Noise. In the 2026 symmetry benchmark, HOG fell from 0.80 to 0.71 skill under pixel noise at σ = 20, while the network readouts absorbed it, widening the deep lead from 0.03 to 0.13.
- Blur. In the same benchmark, an edge-pyramid method (PHOG) collapsed from 0.63 to 0.08 at blur σ = 4 and a fixed Gabor filter bank fell from 0.71 to 0.49.
- Real-world variability. Traditional methods lose accuracy with occlusion, lighting variation or similar-looking objects, for example in pedestrian detection at night.
- Subjective or open-ended defects. Conventional inspection depends on precisely defined, objective defect criteria, whereas human inspectors often apply criteria that are easy to recognise but hard to encode. Deep learning can separate real defects from cosmetic variation without being shown every possible case.
The signal to move to a learned model is variability that explicit rules struggle to capture. If every new batch or supplier needs another threshold, the problem has outgrown explicit rules and a model trained on representative images will usually hold up better. Tyre wear inspection is a good example: lighting, angle, dirt and manufacturing differences change the appearance enough that edge or colour rules struggle. The sibling article on deep learning versus traditional computer vision methods looks at the comparison from the deep-learning side.
Combining classical and deep learning in one production pipeline
MVTec describes a hybrid in which deep learning handles preprocessing or segmentation and classical operators then do the precise geometric measurement or quality evaluation. It also works the other way round: simple classical steps such as resizing, cropping and normalisation can noticeably improve a deep model’s performance and stability. Cognex likewise notes that combining rule-based and AI-powered tools can be the most efficient solution.
The benchmark data hints at why hybrids help. In the 2026 symmetry study, the per-image success correlation between the deep readout and HOG was only 0.46, against 0.78 between the two network-based scorers, so the classical descriptor made different mistakes from the deep one while the two networks largely duplicated each other. Methods that fail differently are exactly what an ensemble or a cross-check stage needs.
A sound split puts a learned model where the input is variable (pre-segmenting the defect region), followed by classical stages that measure and evaluate quality. Each stage can then be logged and tested on its own. A design like this keeps the dimensional result traceable to calibrated geometry even when a network located the region, and it confines model retraining to the one stage that needs it. Our guide to architecting a modular computer vision pipeline covers the interfaces and observability that make this work in production. Designing and building pipelines of this kind, choosing classical, learned or hybrid stages for each step of an inspection, is the work of TechnoLynx’s computer vision R&D team.
What stays open is your own line. The evidence sets defaults, not verdicts. A classical system that looks clean in a controlled trial can still meet lighting drift or a new supplier’s part, and the only reliable test is to run candidate methods on images captured under the conditions the system will actually face.
FAQ
Is computer vision ML or AI?
Computer vision is generally treated as a field of AI, and most modern systems use machine learning models. Classical computer vision is the exception: by definition it relies on engineered features and algorithms instead of machine learning.