The Research Landscape: How the Field Divides

CNNs appear in roughly two-thirds of more than 50 reviewed sources on industrial defect detection — spanning peer-reviewed literature, active patents, and pending applications from institutions across China, Europe, South Korea, Japan, Taiwan, and North America. ViT-based approaches appear in approximately one-fifth of those sources, concentrated in publications from 2021 onward, indicating a rapid but relatively recent surge of interest. A significant and growing segment — roughly one-quarter of the reviewed sources — addresses hybrid CNN-Transformer architectures, suggesting the field is actively moving beyond a binary choice.

Key assignees include institutions in China’s Shandong and Zhengzhou university systems, Karlsruhe Institute of Technology (KIT), University of Wuppertal, University of Udine, Onto Innovation Inc., and Amgen Inc. Dominant technical themes — local vs. global feature extraction, data scarcity, real-time inference constraints, and anomaly detection under limited labeled data — frame every architectural decision engineers must make when selecting between these paradigms.

CNNs: Local Precision and Deployment Maturity

CNNs have been the de facto standard in deep-learning-based computer vision for industrial inspection for over a decade, and their dominance is grounded in concrete performance metrics. A CNN-based system from National University of Tainan (2021) achieved 99.7% accuracy on pump impeller images at 56.87 milliseconds per inference — a throughput directly compatible with live production lines. This combination of accuracy and speed is not incidental; it follows from CNN’s core design around local receptive fields, weight sharing, and hierarchical feature extraction, which efficiently captures texture-level surface anomalies and spatially localised defects.

CNNs also demonstrate versatility across defect types. Research from Kumoh National Institute of Technology (2022) demonstrates CNN applicability in mask manufacturing lines where defects are extremely small in scale and diverse in type. Lightweight CNN variants using depth-separable convolutions — as demonstrated in YOLO-RFF from Qilu University of Technology (2022) — reduce computation while maintaining real-time detection accuracy, making CNNs deployable even in hardware-constrained edge environments. According to IEEE, YOLO-family architectures have become a standard benchmark for real-time object detection in industrial settings precisely because of this efficiency profile.

The principal architectural limitation of CNNs is their reliance on local receptive fields, which constrains their ability to model long-range spatial dependencies across an image. Zhengzhou University’s 2022 study on Swin Transformer combined with CNN explicitly states that “CNNs are more concerned with local information and lack global perception,” making it difficult to extract defect features that span larger spatial regions or vary greatly in scale. This is not a solvable problem through training alone — it is structural, and it is the opening that Vision Transformers were designed to exploit.

“CNNs are more concerned with local information and lack global perception, making it difficult to effectively extract defect features that span larger spatial regions or vary greatly in scale.”

When labeled defect samples are unavailable — a common industrial reality — CNNs shift to anomaly detection frameworks. Research from BIBA, University of Bremen (2019) demonstrates deep metric learning with triplet networks trained exclusively on non-defective surfaces to detect novel anomaly classes not seen during training. This one-class learning paradigm is well-suited to CNNs’ inductive biases, and it remains the dominant practical approach for CNN-based inspection where defect data is scarce.

Vision Transformers: Global Context and Anomaly Localisation

Vision Transformers address the core limitation of CNNs by treating an image as a sequence of fixed-size patches — analogous to tokens in natural language processing — and applying multi-head self-attention mechanisms to model pairwise relationships across all patches simultaneously. This global receptive field from the first layer enables ViTs to capture long-range spatial dependencies that are architecturally inaccessible to standard CNNs, a property directly relevant to defects that span large areas or exhibit irregular spatial distributions.

The earliest industrial application of the base ViT architecture for surface defect detection is illustrated in research from Shanghai Normal University TIANHUA College (2021), which divides defect images into N×N patches and processes them through 12 multi-head attention layers, demonstrating superior classification performance on a steel plate defect dataset. For semiconductor wafer inspection — a domain requiring detection of globally distributed defect patterns — research from Ahsanullah University of Science and Technology (2022) benchmarked four transformer-based models (BEiT, FNet, ViT, and Swin Transformer) on the MixedWM38 wafer defect dataset, with the Swin Transformer emerging as the top performer. According to WIPO‘s technology trend reports, semiconductor inspection is one of the fastest-growing application domains for AI-based visual quality control globally.

ViTs show a particularly clear advantage in patch-level anomaly localisation. The VT-ADL model from the University of Udine (2021) combines patch embedding with a Gaussian mixture density network, leveraging the spatial information preserved by transformer-based patch processing to localise anomalous regions on the MVTec industrial dataset. The Masked Transformer for Image Anomaly Localization, also from the University of Udine (2022), introduces a patch masking mechanism in which each patch is reconstructed only from surrounding context patches — directly avoiding the failure mode where standard autoencoders reconstruct even anomalous content. Both approaches operate with only normal training data, making them applicable to the majority of industrial inspection scenarios where defect samples are unavailable at training time.

The primary barrier to ViT adoption in industry is data. ViTs lack the inductive biases of CNNs — locality and translation equivariance — and typically require either large-scale pre-training or architectural modification to perform well on small industrial datasets. Research from the University of Wuppertal (2022) specifically identifies the perception that ViTs require enormous amounts of data as the core reason their real-world industrial deployment remains sparse despite benchmark superiority. However, this barrier is not insurmountable: integrating multi-layer fully connected structures from VGGNet into a ViT increased classification accuracy by 5.64% on the X-SDD hot-rolled strip steel dataset, demonstrating that architectural adaptation can partially recover ViT performance on small datasets.

Head-to-Head: Four Dimensions That Determine the Right Choice

The choice between CNNs and Vision Transformers for industrial defect detection is not a matter of one architecture being universally superior — it is a function of four specific engineering constraints. Each dimension maps directly to measurable differences in real deployments.

1. Feature Extraction Mechanism

CNNs use spatially local convolutional filters with shared weights, providing translation equivariance and efficient extraction of texture-level features such as scratches, surface irregularities, and localised pits. ViTs divide the image into fixed patches and apply self-attention across all pairs, enabling global spatial reasoning. Zhengzhou University’s 2022 study explicitly quantifies this difference: CNN-based detectors fail to capture global defect context, and the Swin Transformer’s hierarchical windowed attention is proposed as a remedy for defects with large variation in target size.

2. Data Requirements

This is the most significant practical distinction in industrial deployments. CNNs generalise effectively from smaller labeled datasets and are suited to one-class or anomaly detection training regimes. Research from Karlsruhe Institute of Technology (2022) identifies insufficient training data and expensive data generation as primary challenges for deep learning in industrial inspection — challenges that disproportionately affect ViTs due to their need for large-scale pre-training. The University of Wuppertal’s 2022 review specifically notes that sparse real-world application of ViTs is “likely due to the assumption that they require enormous amounts of data to be effective.”

3. Inference Speed and Computational Cost

CNNs — especially lightweight variants using depth-separable convolutions — achieve faster inference and lower memory footprint. Standard ViTs suffer from O(n²) attention complexity with respect to the number of patches, making high-resolution industrial image processing computationally expensive without architectural modifications such as windowed attention (as in the Swin Transformer). For production lines where throughput is non-negotiable, this difference is decisive.

4. Anomaly Localisation Granularity

ViTs show a clear advantage in patch-level anomaly localisation due to their inherent patch tokenisation. The Masked Transformer architecture from the University of Udine demonstrates that self-attention allows anomaly scores to be computed at patch granularity — a natural alignment with pixel-level defect localisation requirements. CNN-based approaches typically require additional upsampling modules or feature pyramid networks to achieve comparable spatial resolution, adding architectural complexity and inference overhead.

The Hybrid Consensus: Where the Field Is Heading

The field is converging on hybrid CNN-Transformer architectures that combine CNN-style local feature extraction with Transformer-based global attention, yielding performance gains over either standalone approach. This is not a theoretical preference — it is reflected directly in active patent filings and peer-reviewed publications from multiple independent research groups.

Qilu University of Technology (2022) proposes teacher-student models combining Swin Transformer with convolutional components in the backbone, explicitly designed to benefit from both local feature extraction and global context modelling. Onto Innovation Inc.’s 2024 patent on substrate defect detection describes a “common backbone” that can be either “deep-convnet-based” or “transformer-based,” enabling task-specific networks for defect detection, die comparison, and anomaly detection to share the same pre-trained feature extractor — a direct recognition that both paradigms serve complementary roles in semiconductor fabrication inspection. According to EPO patent filing trends, semiconductor and electronics manufacturing represents one of the highest-growth domains for AI-based visual inspection IP globally.

A forward-looking signal comes from Qilu Institute of Technology’s 2026 patent on a Self-Supervised Vision Transformer for Real-Time Industrial Defect Detection, which integrates a hybrid Transformer-convolutional backbone trained via self-supervised learning. The patent explicitly notes that CNN models such as ResNet, U-Net, and EfficientNet have a “local receptive field” limitation that constrains their ability to capture global spatial dependencies critical for detecting subtle, irregular, or diffuse defects in high-resolution images — and proposes self-supervised pre-training as the mechanism to address both the global-dependency limitation of CNNs and the annotation-cost limitation of standard ViTs simultaneously.

Fuzhou University holds two active Chinese patents on Transformer-based industrial defect detection using teacher-student training frameworks with self-attention mechanism networks for anomaly detection without labeled defect samples. Amgen Inc. holds patents on image augmentation techniques for automated visual inspection, addressing the data scarcity problem through synthetic image generation — a foundational challenge for both CNN and ViT training pipelines. The convergence of self-supervised learning, synthetic data generation, and hybrid architectures represents the direction the field is taking, as documented across sources reviewed from institutions in China, Europe, South Korea, and North America. Standards bodies including ISO are actively developing frameworks for AI-based industrial inspection quality assurance that will need to accommodate both architecture families.

Qilu University of Technology and Shandong Computer Science Center appear across at least four distinct publications and patents — including YOLO-RFF, AMFF-YOLOX, a defect detection model based on attention and knowledge distillation, and the self-supervised ViT patent — consistently focused on YOLO-family and hybrid Swin-Transformer architectures optimised for real-time industrial use. This concentration of output from a single institution signals a coordinated research programme rather than isolated experiments, and it is a reliable indicator of where production-ready hybrid architectures will emerge first. PatSnap’s innovation intelligence platform, which tracks patent filings and research activity across more than 120 countries, provides the granular assignee-level data needed to monitor these emerging players in real time.