The Patent Landscape: 40+ Filings, Post-2020 Surge

The AI-driven PCB AOI patent landscape encompasses more than 40 active, pending, or recently granted patents and technical disclosures filed between 2004 and 2026 — with the highest concentration of filings occurring after 2020. This acceleration reflects a structural shift: manufacturers are moving away from rule-based optical thresholding toward trained adaptive algorithms capable of generalizing across PCB designs, lighting conditions, and defect morphologies. Dominant assignees include Siemens AG, Tsinghua University, Central South University, Robert Bosch GmbH, Suzhou Kangdai Intelligent Technology, and United Automotive Electronics, with filings spanning China, Japan, Korea, and the United States.

The dominant technical approaches across the dataset fall into four categories: (1) transfer learning and fine-tuning of pre-trained vision models — CNNs, Vision Transformers, and CLIP-based architectures — onto PCB-specific defect datasets; (2) hybrid AOI pipelines where optical image processing provides initial defect flags that a trained adaptive algorithm refines or overrides; (3) cross-modal prompt learning combining visual and text encoders; and (4) continual learning for incremental defect type expansion. A secondary but significant cluster covers YOLO-family model adaptations for real-time PCB defect detection, with multiple Chinese university assignees reporting improvements in mAP, recall, and false-negative rate through structural modifications combined with fine-tuning on PCB-specific datasets.

Hybrid AOI Pipelines: Foundation Models as Second-Stage Classifiers

The most architecturally significant innovation in recent PCB AOI patents is the replacement — or augmentation — of classical rule-based optical image processing with a trained adaptive algorithm acting as a second-stage decision maker. Rather than relying solely on threshold-based comparisons, hybrid pipelines use a pre-trained machine learning model to re-evaluate candidate defects flagged by the optical subsystem, directly addressing the chronic problem of false positives that previously required operator-intensive manual review and test-program maintenance after every process or lighting change.

Siemens AG has been the most prolific patent filer in this architectural category. The Method for Optical Quality Control During Circuit Board Manufacturing (Siemens, 2023) defines a workflow where a first defect indicator is produced by conventional optical image processing, and only when this indicator flags a defect is a second defect indicator produced by a trained adaptive algorithm. The output of the ML model is then substituted for the optical result. A complementary Siemens patent, Inspection of Printed Circuit Board Assemblies (Siemens, 2024), further specifies that AOI numerical measurement results — extracted as feature vectors with dimensionality reduction — can be fed into a classifier trained on image-label pairs, with the approach explicitly targeting false-call detection in SMT manufacturing.

“A shared Visual Foundation Model generates a universal feature representation, while multiple task-specific detection sub-models are fine-tuned per production line domain — making fine-tuning both targeted and computationally efficient.”

Robert Bosch’s approach, disclosed in Method and Apparatus for Automated Optical Inspection (Robert Bosch, 2025), introduces a domain-specific decomposition: a shared visual foundation model — described as a hierarchical Vision Transformer backbone using architectures such as Swin, CvT, CSwin, SAM, or DINOv2 — generates a universal feature representation, while multiple task-specific detection sub-models are fine-tuned per production line domain. The patent explicitly distinguishes between a “general feature extraction sub-model” that is pre-trained and frozen, and task detection sub-models that are domain-adapted. Each domain-specific sub-model is trained on images and labels collected from its specific production environment, making fine-tuning both targeted and computationally efficient. This represents the most explicit treatment of foundation model fine-tuning in the entire dataset, as documented by EPO filing records.

The X-ray inspection domain follows the same principle. The System and Method for Automatic X-Ray Inspection (Siemens, 2023) trains 3D convolutional neural networks on ground-truth labeled X-ray image stacks — including cross-sectional slice sequences — to classify solder-joint defects such as pin shorts. The CNN is supervised with labeled image sets annotated with defect categories, producing fine-tuning data that enables the model to generalize across solder joint geometries. This approach aligns with broader trends in industrial computer vision documented by IEEE and WIPO in their respective technology and patent trend reports.

Cross-Modal Prompt Learning and Vision-Language Fine-Tuning

Vision-language foundation models — architectures pre-trained on paired image-text data — can be fine-tuned for PCB anomaly detection through prompt engineering and cross-modal alignment, achieving accurate defect localization with minimal labeled samples. This approach is among the most technically advanced in the dataset, emerging exclusively from 2025 filings.

Central South University filed two closely related patents in 2025 covering this paradigm. The PCB Defect Detection Method and System Based on Cross-Modal Prompt Learning and Visual Guidance constructs a model with a pre-trained visual encoder and a pre-trained text encoder as the feature extraction backbone. Normal and anomalous PCB states are encoded as text prompts, and the model learns a joint cross-modal feature space. During fine-tuning, a dual-constraint alignment strategy combining contrastive loss and triplet loss adjusts the mapping between visual and text features to establish a clear normal/anomaly boundary.

The system also introduces a dual-channel repository mechanism with hard-negative sample mining: synthetic anomaly samples generated from defect-free PCB images using structure-and-physics-prior augmentation are screened to select the most training-informative examples. The inference pipeline compares the visual encoder’s spatial feature vectors against stored text feature embeddings at every spatial position to generate similarity maps, attention maps, or heat maps, which are then upsampled and thresholded to produce binary defect localization masks. This approach is explicitly claimed to improve accuracy, robustness, and generalization under data-sparse conditions — a key bottleneck in PCB AOI given the rarity of real defect samples.

Zhejiang University extends cross-modal reasoning to a PCB-specific anomaly detection network, PCNet, in PCB Multi-Level, Multi-Scale Anomaly Detection Network PCNet (Zhejiang University, 2025). The method inputs PCB images into a frozen CLIP model’s visual encoder to obtain initial visual embedding feature maps, then passes them through a series of prompt convolution blocks that effectively fine-tune the response of the frozen foundation model without full retraining. Synthetic anomalous images with labels are generated from normal samples and used as training data, producing a system capable of effective anomaly detection with only a small number of normal training examples.

Transfer Learning, Continual Learning, and Few-Shot Adaptation

A large cluster of patents addresses the practical challenge of limited labeled defect data by combining pre-trained model fine-tuning with data augmentation, active learning, and continual learning strategies — each targeting a different bottleneck in the AOI deployment lifecycle.

Continual Learning to Prevent Catastrophic Forgetting

Tsinghua University’s PCB Defect Detection Method Based on Continual Learning (2023) is particularly significant: it addresses catastrophic forgetting in deployed AOI models. Using a YoloX backbone with a CSPDarknet trunk, FPN feature pyramid, and three YoloHead detectors, the model is designed to incrementally learn new defect categories from new production samples without forgetting previously learned categories. When the model encounters a defect type it was not trained on, it uses the new image to update itself — effectively implementing online fine-tuning in a manufacturing deployment setting. This capability is critical when PCB designs or manufacturing processes change and new defect signatures emerge.

Active Learning and Semi-Supervised Domain Adaptation

United Automotive Electronics filed two closely related patents (2021; 2024) describing a re-inspection model fine-tuned by deep learning algorithms on AOI output files. The model selects the correct re-inspection sub-model based on the imaging algorithm associated with each PCB image type, outputs confidence scores and heat maps, and uses an active learning strategy to optimize the model on partially labeled samples — directly reducing the human verification workload while continuously improving detection accuracy. Maximum Mean Discrepancy (MMD) and KL divergence are used as distribution metrics for domain adaptation between labeled and unlabeled samples, a form of semi-supervised fine-tuning.

Few-Shot Prototype Learning

The University of Science and Technology of China addresses few-shot fine-tuning in PCB Defect Detection Method Based on Few-Shot Learning (USTC, 2023). An instance-level embedding backbone is trained with position-weight assignment modules and sample-weight assignment modules — effectively learning defect-class-aware prototypes from only a few labeled examples. Cross-entropy loss fine-tunes the backbone, while defect class support sets are used to compute attention prototypes, enabling recognition of rare defect categories without large-scale retraining. This approach is particularly relevant given the rarity of real defect samples in production environments, a challenge well-documented by Nature in machine learning for manufacturing research.

Jiangxi Huashi Optoelectronics’ AOI Defect Detection Method and System Based on Image Processing (2026) explicitly integrates transfer learning and meta-learning into the fine-tuning pipeline. A conditional GAN generates defect-free reference images as comparison baselines, while spatial attention mechanisms and multi-scale feature pyramid networks are combined with a pre-trained defect classification model that is fine-tuned using transfer learning and meta-learning frameworks. The patent claims this combination significantly reduces false-positive and missed-detection rates while improving adaptability and generalization across complex texture backgrounds.

Key Players: Industrial OEMs vs. Research Institutions

The patent data reveals a clear bifurcation between industrial OEM players who focus on production-ready hybrid pipelines and research institutions that advance the algorithmic frontiers of fine-tuning. Each group is solving a different version of the same problem: how to deploy adaptive ML models in manufacturing environments where labeled data is scarce, defect types evolve, and false-positive rates have direct cost consequences.

Siemens AG is the dominant industrial assignee with at least three major PCB/PCBA AOI patents (2023–2024), all filed in China and all centered on the trained adaptive algorithm as a second-stage decision-maker layered on top of classical optical inspection. Siemens’ IP consistently targets the full manufacturing loop — from SMT to X-ray to PCBA-level feature vector classification. Robert Bosch GmbH contributes the most architecturally explicit treatment of foundation model fine-tuning (2025), distinguishing between frozen pre-trained feature extractors (Swin, DINOv2, SAM) and domain-fine-tuned task sub-models across multiple production lines.

Tsinghua University is the dominant Chinese academic assignee with at least four patents (2020–2023) covering CNN-based detection, real-time cloud deployment, and continual learning. The cloud-connected architecture in the 2020 patent allows centralized model updates without local re-programming — a deployment model that anticipates modern MLOps practices. Central South University leads in cross-modal and vision-language fine-tuning, with two 2025 patents on CLIP-based prompt learning for PCB anomaly detection. Suzhou Kangdai Intelligent Technology implements a cascaded multi-sub-model architecture where each sub-model is pre-trained on a specific defect type, and mismatched predictions trigger human annotation followed by retraining — implementing a supervised fine-tuning feedback loop directly in the production environment.

IBM introduced a specialized visual inspection training board concept (2024), where physical PCBs with intentionally embedded defects are used as dedicated training artifacts for fine-tuning AI algorithms in the AOI system — a hardware-assisted fine-tuning approach that removes dependence on rare production defect samples. YOLO-family model fine-tuning is the most frequently cited detection architecture across smaller Chinese university assignees — including Shanghai University, HDUET, Huaiyin Institute of Technology, Northwest Polytechnical University, and Guangzhou Railway Polytechnic — all reporting improvements in mAP, recall, and false-negative rate through structural modifications such as attention modules, multi-scale feature pyramids, and improved IoU loss functions combined with fine-tuning on PCB-specific datasets. These trends are consistent with broader industrial AI adoption patterns tracked by OECD in its manufacturing technology reports.