Computer Vision Defect Detection Patents 2026
Computer Vision Defect Detection Patents 2026
The field is at an inflection point in 2026, driven by transformer-based attention, event-camera sensing, and edge-deployable AI. This report characterizes innovation signals across 50+ retrieved patent and literature records spanning 2016–2026.
Deep Learning Transforms Industrial Defect Detection
Computer vision defect detection divides into five technical sub-domains in this dataset: deep learning model architectures, multi-modal and multi-spectral imaging hardware, training data strategies for low-sample regimes, unsupervised and generative model approaches, and edge/real-time deployment optimization. The retrieved dataset spans 17 distinct assignees across CN, US, WO, EP, TW, and IN jurisdictions.
Core mechanisms across the dataset include CNNs applied to surface image data, GANs and reconstruction-based approaches for anomaly detection without labeled defect samples, multi-scale feature pyramid architectures for small-defect localization, and attention mechanisms for feature fusion across imaging modalities. Signal-to-noise ratio quantification and nuisance filtering are increasingly formalized within semiconductor inspection pipelines.
The mid-phase cluster from 2019–2022 is anchored by KLA Corporation’s multi-jurisdictional neural network training portfolio for semiconductor wafer inspection, alongside Accenture’s ML-based intelligent defect detection platform (2021), Fujitsu’s GAN-based unsupervised defect detection (2021–2023), and Mitutoyo’s training-volume guidance system (2021). At least 10 review and methodology papers published between 2020 and 2023 reflect rapid consolidation of deep learning as the dominant paradigm.
The most recent filing cluster (2024–2026) signals significant architectural evolution: event-camera fusion, multi-modal encoder-decoder pipelines, knowledge distillation for edge hardware, 360° robotic surface capture, and borescope-based 3D depth estimation. In this dataset, KLA Corporation accounts for 9 of the 35 retrieved patent records, the highest filing count among all assignees in retrieved records.
Jurisdiction and Technology Cluster Breakdown
Among the 35 retrieved patent records, US filings (19 records) dominate semiconductor, aerospace, and automotive inspection, while CN filings (9 records) concentrate on manufacturing-floor applications, event camera integration, and edge-deployment architectures.
Patent Records by Jurisdiction — Computer Vision Defect Detection (Dataset Snapshot)
In this dataset, the US accounts for 19 of 35 patent records, followed by CN (9), WO (7), TW (2), EP (1), and IN (1), reflecting the concentration of semiconductor and aerospace inspection IP among US-based assignees in retrieved records.
↗ Click bars to explorePatent Filing Activity by Period — Computer Vision Defect Detection (Dataset Snapshot)
In this dataset, filing activity accelerated sharply from the 2019–2022 period onward, with the 2023–2026 cluster yielding the highest concentration of architecturally novel filings including event-camera fusion, multi-modal encoders, and SNR-guided model tuning in retrieved records.
↗ Click bars to exploreKey Inspection Domains Covered in Retrieved Records
The retrieved dataset reveals distinct IP clusters across semiconductor wafer inspection, automotive and railroad inspection, aerospace turbine maintenance, and consumer electronics surface QC, each with domain-specific sensor and model architectures.
Semiconductor Wafer Inspection
KLA Corporation (9 retrieved records) and Applied Materials Israel Ltd. (2 records) focus entirely on wafer, reticle, and flat panel inspection across US, WO, CN, and TW jurisdictions (2019–2026). KLA’s cascaded nuisance filtering architecture uses a first network to broadly filter defect candidates and a second high-resolution network to refine results. KLA’s 2026 filing introduces a per-location signal-to-noise metric for annotation guidance and model robustness quantification. Nanotronics Imaging’s active-learning system iteratively labels images until a training threshold is reached for nanoscale surface inspection.
Semiconductor ManufacturingAutomotive and Railroad Inspection
Hyundai Mobis Co., Ltd. filed 2 US patents (2023, 2025) covering a deep learning vision inspection system for automotive parts with a ground-truth generation module and learning module. Transportation IP Holdings, LLC filed 2 US patents (2025, 2026) on a two-stage machine learning pipeline for railroad vehicle underbody inspection, incorporating area-of-interest extraction followed by defect-type and severity classification with leak visibility enhancement. Both assignees target safety-critical part inspection in high-throughput production contexts.
Automotive ManufacturingAerospace and Energy Infrastructure
RTX Corporation’s 2025 US and WO patents describe a borescope-based system reconstructing 3D defect geometry from 2D borescope images for turbine engine inspection, funded in part by a US Air Force government contract. Baker Hughes Holdings LLC (2022, WO) and Baker Hughes Oilfield Operations LLC (2026, US) apply a cascaded two-algorithm computer vision pipeline for non-destructive testing of pipeline and oilfield equipment, where a fast first algorithm gates access to a slower, more accurate second algorithm. FARO Technologies’ 2021 US patent covers AI-based construction defect detection on a mobile scanning platform correlated with 2D spatial maps.
Aerospace and Energy NDTConsumer Electronics Surface QC
Jiangsu Techuang Technology Co., Ltd. filed 2 CN patents (2025, 2026) covering a 6-axis robotic arm system for 360° surface capture of laptop computers, feeding a deep learning model with a dynamic classification threshold incorporating IoU, texture entropy, and edge density. Hangzhou Qijing Technology Co., Ltd.’s 2026 CN patent describes an intelligent detection system using structured light source modules, array industrial cameras, spectral imaging, and knowledge distillation for deployment on high-parallel hardware accelerators with real-time integration into manufacturing execution systems. Mitutoyo Corporation’s 2021 US patent provides active guidance on the number of additional defect images required for training with metrological measurement capability.
Consumer Electronics QCLeading Patent Assignees in Computer Vision Defect Detection — Dataset Snapshot
In this dataset, KLA Corporation / KLA-Tencor Corporation accounts for 9 of 35 retrieved patent records — the highest filing count among all assignees in retrieved records — spanning US, WO, CN, and TW jurisdictions. A second tier of assignees each contributes 2 retrieved records, including Fujitsu Limited, Applied Materials Israel Ltd., Nanotronics Imaging, Hyundai Mobis, Transportation IP Holdings, RTX Corporation, and Baker Hughes.
Top Assignees by Filing Count in Retrieved Records (Dataset Snapshot)
↗ Click bars to exploreKLA Corporation
KLA Corporation / KLA-Tencor Corporation holds 9 retrieved records in this dataset — the highest count among all assignees — spanning US, WO, CN, and TW jurisdictions across filings dated 2019–2026. Key technology areas include paired high-resolution/low-resolution neural network training for wafer inspection, cascaded deep learning nuisance filtering to reduce false positives, and a 2026 signal-to-noise metric patent (US/WO) for annotation guidance and model robustness. The portfolio is entirely focused on semiconductor wafer, reticle, and flat panel inspection and represents a multi-jurisdictional IP position.
United StatesFujitsu Limited
Fujitsu Limited holds 2 US patents in this dataset (2021, 2023) covering GAN-based unsupervised defect detection where a generator neural network is trained exclusively on defect-free samples, synthetic defects are superimposed during training, and live images are compared against defect-free reconstructions at inference to localize anomalies. This approach directly addresses the labeled-data scarcity problem and represents a commercially significant alternative to supervised pipelines. Both patents are US-granted filings targeting industrial surface inspection without the need for labeled defect corpora.
JapanFive Frontier Directions in 2024–2026 Filings
Among the most recent filings in this dataset (2024–2026), five distinct architectural and methodological directions are visible, ranging from neuromorphic event camera fusion to quantitative 3D defect characterization and SNR-guided model management.
Event Camera + Conventional Camera Fusion
Pazhou Lab (Huangpu)’s 2024 CN filing introduces neuromorphic event cameras combined with conventional cameras in a stereo configuration using CSPDarknet as the backbone and a cross-attention mechanism to fuse feature maps from both sensors. This architecture enables high-performance defect detection under complex or rapidly changing illumination conditions. The cross-attention feature fusion applied to event frames and RGB frames represents a frontier not previously seen in earlier filings in this dataset.
Knowledge Distillation for Edge Deployment
Hangzhou Qijing Technology Co., Ltd.’s 2026 CN filing explicitly describes knowledge distillation to compress multi-modal inspection models for deployment on high-parallel hardware accelerators, with real-time integration into manufacturing execution systems. The system combines structured light source modules, array industrial cameras, and spectral imaging with attention-based feature extraction. This represents a direct response to the need to run architecturally complex models on manufacturing-floor edge hardware without sacrificing detection capability.
Supervised vs. Unsupervised Defect Detection: Key Dimensions
Click any row to explore further.
| Dimension | Supervised Deep Learning | Unsupervised / Generative Approaches |
|---|---|---|
| Training data requirement | Requires labeled defect image corpus; volume guidance provided by systems like Mitutoyo’s 2021 US patent | Trains on defect-free samples only; synthetic defects superimposed (Fujitsu, 2021–2023 US patents) |
| Representative assignees in this dataset | KLA Corporation, Hyundai Mobis, Nanotronics Imaging, Accenture (active-learning labeling loop) | Fujitsu Limited, Applied Materials Israel Ltd., Accenture (augmentation-based synthesis) |
| Detection capability | High accuracy on known defect classes; Nanotronics uses iterative active-learning until training threshold reached | Detects both known defect classes and novel anomalies; Applied Materials combines supervised and unsupervised outputs via optimized weighting |
| False positive handling | KLA’s cascaded nuisance filtering uses two-stage deep learning networks to reduce false alarms without sacrificing recall | Anomaly score thresholding at inference; reconstruction error used to localize anomalies (Fujitsu approach) |
| Edge / deployment readiness | Knowledge distillation applied to compress supervised models for high-parallel hardware (Hangzhou Qijing, 2026 CN) | Reconstruction-based inference is computationally intensive; edge deployment less formalized in retrieved records |
| Primary application domains | Semiconductor wafer inspection, automotive parts, railroad underbody, aerospace turbines | Industrial surface inspection, lens inspection, general manufacturing anomaly detection |
| Metric innovation | KLA’s 2026 US/WO SNR metric formalizes per-location annotation guidance and model robustness quantification | Composite grade map combining supervised and unsupervised model outputs (Applied Materials Israel, 2021 US) |
Frequently Asked Questions: Computer Vision Defect Detection Patents
KLA Corporation / KLA-Tencor Corporation holds 9 of the 35 retrieved patent records in this dataset — the highest filing count among all assignees — spanning US, WO, CN, and TW jurisdictions with filings dated 2019–2026, entirely focused on semiconductor wafer, reticle, and flat panel inspection.
KLA’s foundational patent family pairs a high-resolution neural network with a low-resolution inspection network. The high-resolution network synthetically generates defect images, which are then used to train the low-resolution network, enabling high-throughput wafer inspection without requiring a large labeled low-resolution defect corpus.
Fujitsu’s 2021 and 2023 US patents describe a generator neural network trained exclusively on defect-free samples. Synthetic defects are superimposed during training to build reconstruction capability, and at inference live images are compared against the defect-free reconstruction to localize anomalies — requiring no labeled defect images.
The 2024–2026 filings include: event camera plus conventional camera fusion with cross-attention (Pazhou Lab, 2024 CN); multi-modal RGB, polarized light, and interferometric encoder-decoder architectures for lens inspection (Shenzhen Tianding, 2025 CN); knowledge distillation for edge hardware deployment (Hangzhou Qijing, 2026 CN); 3D borescope depth estimation for turbine inspection (RTX Corporation, 2025 US/WO); and a per-location signal-to-noise metric for annotation guidance (KLA, 2026 US/WO).
US filings account for 19 of 35 retrieved patent records, dominated by semiconductor inspection (KLA, Applied Materials), aerospace (RTX), and automotive (Hyundai Mobis). CN filings number 9 records and concentrate on manufacturing-floor applications, event camera integration, and edge-deployment architectures from assignees including Pazhou Lab, Shenzhen Tianding, Hangzhou Qijing, and Jiangsu Techuang.
RTX Corporation’s 2025 US and WO patents describe a system that reconstructs 3D defect geometry from 2D borescope images for turbine engine inspection, moving beyond binary defect localization toward quantitative depth characterization. The work was supported in part by a US Air Force government contract and is relevant to aerospace maintenance and remaining-useful-life assessments.
Data and insights on this page are based on a limited patent and literature dataset and are for reference only. Figures may not represent the complete technology landscape.