How rule-based machine vision encodes defect knowledge — and where it breaks down

Traditional rule-based machine vision systems detect surface defects by comparing captured images against pre-defined statistical or photometric criteria — every inspection criterion must be manually encoded by an expert before the system is deployed. A representative example is the variance-of-variance algorithm described by Pusan National University (2011), which divides the inspection surface into a grid of fractional areas and applies a probability-based mathematical algorithm derived from the luminance distribution of each sub-region. The system identifies defects including scratches and paint faults by detecting statistical anomalies in brightness across these fixed local windows.

Konica Minolta’s workpiece surface defect detection system (2020) illustrates a more sophisticated rule-based approach: a moving light-dark pattern is projected relative to the workpiece and multiple images are acquired. Defect candidates are extracted by thresholding the number of images in which a provisional defect appears — if this count meets a preset minimum, the candidate is confirmed as a true defect. The confirmed candidates are then composited into a single image for final detection. This structured illumination approach improves sensitivity to small surface defects but still depends entirely on pre-set numerical thresholds, meaning it cannot generalise to defect types not anticipated during system configuration.

Semiconductor Technologies & Instruments’ wafer inspection system (2012) exemplifies a statistical reference-image approach: a training process captures multiple images of a reference wafer, calculates statistical parameters such as mean and variance per pixel, and uses these to construct a plurality of reference images. Inspection images are then compared against this statistical model. This pixel-by-pixel statistical comparison is rigorous but brittle — it presupposes stable imaging conditions and is largely incapable of classifying the type or severity of a defect beyond pass/fail thresholding. According to ISO quality management standards, deterministic threshold systems remain valued for their auditability in regulated manufacturing environments, but their inability to generalise represents a structural ceiling on classification capability.

Raytheon’s automatic sight-check system (1993, Japan) used paired cameras under different illumination conditions — bright-field and dark-field — with rule-based image attribute comparison to inspect electronic modules. The system compensated for non-uniform surfaces by comparing image models under geometrically controlled lighting. While innovative for its era, this approach required extensive manual calibration per product type, a limitation that scales poorly as product variety increases. The fundamental constraint of all rule-based systems is the same: defect knowledge must be fully specified in advance, expressed as measurable photometric parameters, and re-engineered whenever a new defect type emerges.

What AI architectures bring to surface defect classification

AI-powered inspection systems fundamentally depart from rule-based methods by learning feature representations directly from labeled or unlabeled image data — eliminating the need to manually specify every defect criterion. The most prevalent AI architecture in the patent data is the convolutional neural network (CNN). CooperVision International Limited’s system (2023) preprocesses ophthalmic lens images and inputs them into a CNN that analyses and characterises defect types, identifies defect regions, and outputs defect categories or classifications for each image. Separate “edge” and “surface” AI models specialise classification by the part of the lens being inspected — a degree of contextual differentiation impossible in threshold-based systems.

The defect classification hierarchy achievable with AI is significantly more granular than rule-based approaches. CooperVision’s 2024 patent family describes a system capable of outputting classifications across multiple lens surface defect classes and lens edge defect classes — including good lens, bubble, scratch, edge split, squeezed lens, edge notch, lens too small, lens too large, and lens too eccentric. Class Activation Maps (CAMs) are employed to visualise which spatial regions drove the network’s classification decision, providing interpretability alongside accuracy. Research published via Nature has consistently demonstrated that CNN-based visual recognition outperforms hand-crafted feature extraction on complex texture discrimination tasks, which is precisely the challenge surface defect classification presents.

“AI systems extract defect knowledge from data rather than requiring an expert to manually encode every inspection criterion — enabling classification of anomalies that were never anticipated during system design.”

Unsupervised and generative AI approaches extend classification capability to defect types not explicitly labeled during training. Daekhon International’s system (2018) divides images into blocks, compresses each block into a feature vector via a convolutional autoencoder, reconstructs the block, and identifies defects by measuring reconstruction error. Because the system is trained only on normal product images, it can flag anomalies without requiring labeled defect examples — a key advantage in manufacturing environments where defects are rare or diverse. Beijing Baidu Netcom Science and Technology’s GAN-based approach (2021) trains a GAN exclusively on defect-free surface images: when a surface image flagged as defective by a primary detection model is passed into the GAN, the GAN’s ability or inability to reconstruct it as a normal surface determines whether the defect detection result is confirmed or revised.

Mitsubishi Electric Corporation’s domain-conversion approach (2021) transforms a target image through sequential forward and backward domain conversions trained on normal images, then compares the result to the original target image to reveal anomalies. NEC Laboratories America’s system (2021) generates classification weights from both visual features and semantic descriptions, enabling few-shot or zero-shot defect classification without retraining — a capability completely absent from rule-based systems. According to IEEE standards bodies, zero-shot learning approaches represent a significant frontier in industrial AI, particularly for high-mix manufacturing where new product variants arrive continuously.

Industry applications: pharmaceuticals, steel, electronics, and beyond

The patent data reveals that AI-powered visual inspection has been adopted across a broad range of manufacturing sectors, each presenting distinct defect classification challenges that expose the limitations of purely rule-based approaches. The application domains differ markedly in their defect rarity, product variety, and regulatory requirements — factors that directly shape which inspection architecture is selected.

Pharmaceutical and biomedical manufacturing

Amgen Inc. is the most prolific assignee in this dataset, with at least seven distinct filings across six jurisdictions covering automated visual inspection (AVI) for pharmaceutical containers. Their image augmentation approach addresses a core limitation of AI-based systems: the need for large, balanced training datasets when defects are rare in production. Amgen’s arithmetic transposition algorithm transposes defect features from one image onto the background of a clean image at pixel-level realism, generating synthetic training data. Additionally, digital inpainting using deep learning models can add, remove, or modify depicted defects in images. A parallel Amgen deep learning platform patent (2023) describes automatic downsampling and selective cropping of container images to reduce the computational cost of training and inference while preserving classification accuracy.

Steel and industrial manufacturing

JFE Steel Corporation’s inspection method for steel surfaces (Brazil, 2024) demonstrates a two-stage AI pipeline: candidate defect parts are first extracted from the captured surface image, then screened by a first defect determination step, and finally subjected to a detailed inspection step using a trained model. This staged architecture improves processing time by only applying computationally expensive AI inference to pre-screened regions of interest. Hyundai Steel’s surface inspection system (Korea, 2025) uses a real-time AI classification model combined with a worker-history-based self-learning model that can detect whether a flagged defect represents a new, previously unseen defect type and recommend it for inclusion in the training dataset — a capability with no analogue in rule-based systems.

Electronics and component inspection

Tyco Electronics (Shanghai)’s automated part inspection system (2024) explicitly runs a machine vision inspection module and an AI inspection module in parallel on the same digital image, comparing both outputs to generate a final quality determination. This design acknowledges that rule-based machine vision retains advantages in detecting specific, well-characterised dimensional deviations, while AI excels at classifying ambiguous or complex texture-based defects. Samsung Electronics holds patents (2019, 2022) covering vision inspection management methods that integrate process data — linking inspection outcomes to upstream manufacturing parameters in ways that purely image-based rule systems cannot.

Process management and data quality

Panasonic’s data gathering conditions checking system (US, 2022) addresses a critical infrastructure concern unique to AI systems: the model’s sensitivity to data capture condition drift. By generating representative images for image groups and computing luminance-based difference indices, the system detects when imaging conditions — lighting, focus, camera alignment — have shifted enough to invalidate AI model outputs. This quality assurance layer has no equivalent in purely rule-based systems, which are directly parameterised by illumination conditions rather than implicitly sensitive to them. The WIPO patent database records growing filings in this monitoring category, reflecting industry recognition that AI inspection infrastructure requires active maintenance in ways that rule-based systems do not.

Head-to-head: classification granularity, false positives, and speed

The most significant architectural contrast between AI-powered visual inspection and rule-based machine vision lies in how defect knowledge is encoded and how that encoding determines system performance across seven key dimensions. The comparison below is drawn directly from the patent evidence.

False-positive reduction is a shared challenge across both paradigms. In AI systems, Baidu’s two-stage architecture uses a GAN trained on defect-free images to confirm or revise detections from a primary detection model. Musashi AI North America’s hybrid system (2024) requires both an AI object detection model and a golden sample comparison step to agree before confirming a defect. Rule-based systems address false positives via composite multi-image threshold confirmation, as in Konica Minolta’s structured illumination system (2020), where a defect candidate must appear in a minimum number of images before confirmation.

Processing speed is another dimension where rule-based systems retain a structural advantage. Simple arithmetic operations over fixed image regions execute faster than deep neural network inference. However, JFE Steel’s two-stage AI pipeline (2024) mitigates this by restricting expensive AI inference to pre-screened regions of interest, and Amgen’s downsampling approach (2023) reduces the computational cost of both training and inference while preserving classification accuracy. The latency gap between the two approaches is narrowing as inference hardware — particularly GPU and NPU accelerators — continues to improve.

Why hybrid AI + rule-based architectures are the leading engineering trend

The patent evidence points clearly toward hybrid architectures — combining AI detection with rule-based dimensional verification — as the dominant engineering direction in surface defect inspection. These systems are not a compromise between two paradigms; they are a deliberate exploitation of the complementary strengths of each.

OPTOCOMB (Japan, 2025) represents an emerging trend toward hybrid 2D AI + 3D rule-based defect sizing: AI inference identifies defect regions in 2D images and rule-based processing quantifies defect geometry from corresponding 3D shape data. This pipeline reflects a principled division of labour — AI handles the perceptually complex task of identifying where and what a defect is, while rule-based logic handles the metrologically precise task of measuring how large it is. Tyco Electronics (Shanghai)’s 2024 system runs a machine vision inspection module and an AI inspection module in parallel on the same digital image, comparing both outputs for a final quality determination, explicitly acknowledging that neither approach alone is sufficient.

Playoni’s defect inspection method (Korea, 2024) chains three specialised AI models — one for anomaly detection, one for type classification, and one for attribute estimation — before applying final rule-based pass/fail criteria per attribute. This architecture reflects the convergent trend most clearly: AI handles the perceptually complex tasks, while rule-based logic governs the final, auditable decision boundary. Musashi AI North America’s multi-model architecture (2023) escalates ambiguous cases from a first neural network to a second specialised model under defined conditions, improving recall on difficult defect types without increasing false positives on clear cases.

“The convergent trend is clear: AI handles the perceptually complex tasks of defect identification and classification, while rule-based logic governs the final, auditable decision boundary — combining the strengths of both paradigms.”

The emergence of hybrid architectures also reflects practical regulatory and auditability requirements. In pharmaceutical manufacturing — where Amgen operates — every inspection decision must be traceable and defensible to regulators. Pure AI systems that output a classification without a direct parameter mapping are difficult to validate under frameworks such as those published by the FDA for automated inspection in drug manufacturing. Hybrid architectures that use AI for classification and rule-based logic for the final accept/reject boundary provide the interpretability required for regulatory submission while retaining the classification breadth of AI.

NEC Laboratories America’s few-shot and zero-shot classification approach (2021) — generating classification weights from both visual features and semantic descriptions — points toward a future where hybrid systems can be reconfigured for new defect types without full retraining, simply by providing a semantic description of the new defect class. This capability, combined with the dimensional precision of rule-based 3D measurement, represents the frontier of surface defect inspection architecture as reflected in the most recent filings in this dataset. PatSnap’s innovation intelligence platform tracks over 2 billion data points across global patent filings, enabling R&D teams to monitor exactly these convergent trends as they emerge.