Application Research of a Deep Learning Model Integrating CycleGAN and YOLO in PCB Infrared Defect Detection

Chao Yang; Haoyuan Zheng; Yue Ma

arxiv: 2601.00237 · v2 · submitted 2026-01-01 · 💻 cs.CV · cs.LG· cs.RO

Application Research of a Deep Learning Model Integrating CycleGAN and YOLO in PCB Infrared Defect Detection

Chao Yang , Haoyuan Zheng , Yue Ma This is my paper

Pith reviewed 2026-05-16 18:27 UTC · model grok-4.3

classification 💻 cs.CV cs.LGcs.RO

keywords PCB defect detectionCycleGANYOLOv8infrared imagingdata augmentationimage-to-image translationlow-data learningindustrial inspection

0 comments

The pith

CycleGAN translates visible PCB images into synthetic infrared ones to train YOLOv8 detectors that nearly match fully supervised performance despite scarce real IR data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes a framework that uses CycleGAN to map abundant visible-light PCB photographs into the infrared domain through unpaired translation. The resulting pseudo-IR images retain defect shapes and thermal patterns, then get mixed with a small number of real infrared samples to train a lightweight YOLOv8 detector. Experiments show this mixed training set improves detection accuracy over models that use only the limited real data and reaches close to the results obtained from large fully labeled IR datasets. The method therefore offers a practical route around the high cost of collecting and annotating infrared defect images for industrial inspection tasks.

Core claim

By performing unpaired image-to-image translation with CycleGAN, the method produces high-fidelity pseudo-IR PCB samples that preserve defect structural semantics and simulate thermal distributions; fusing these generated samples with limited real IR images then trains a YOLOv8 detector whose performance surpasses training on real data alone and approaches that of fully supervised models.

What carries the argument

CycleGAN unpaired visible-to-infrared translation that generates pseudo-IR training samples for a YOLOv8 defect detector.

If this is right

The detector achieves higher defect detection accuracy under infrared imaging when real training data is limited.
Pseudo-IR synthesis provides a scalable augmentation strategy for industrial vision systems.
Performance nears that of models trained on abundant real infrared samples.
The approach lowers the expense of gathering and labeling large real IR defect datasets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same unpaired translation tactic could transfer to other domains where visible images are plentiful but infrared or thermal data remain scarce.
Refining the translation step to capture finer thermal gradients might further narrow the remaining gap to full supervision.
Applying the pipeline to additional electronic components or defect classes would test its generality beyond standard PCBs.

Load-bearing premise

CycleGAN can translate visible PCB images to infrared while keeping defect structures intact and producing realistic thermal patterns.

What would settle it

If a YOLOv8 model trained on the mixed pseudo-plus-real set shows no accuracy gain over one trained only on the limited real IR data, or if generated images visibly distort defect locations or heat signatures, the central claim would be falsified.

Figures

Figures reproduced from arXiv: 2601.00237 by Chao Yang, Haoyuan Zheng, Yue Ma.

**Figure 1.** Figure 1: Research Roadmap size pseudo-infrared defect images from visible images via unpaired image-to-image translation, preserving defectrelevant structures without requiring paired training data. The generated pseudo-IR images are then combined with real infrared samples to construct an augmented dataset, which is used to train a YOLO-based detector for defect identification and localization, with performance v… view at source ↗

**Figure 3.** Figure 3: a b:Visible light image; c d:Model-generated infrared im [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: Comparison of pseudo-infrared images generated (a) and [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 6.** Figure 6: Training results on real infrared defect images [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗

**Figure 8.** Figure 8: Changes in detection results of the mixed dataset at differ [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗

read the original abstract

This paper addresses the critical bottleneck of infrared (IR) data scarcity in Printed Circuit Board (PCB) defect detection by proposing a cross-modal data augmentation framework integrating CycleGAN and YOLOv8. Unlike conventional methods relying on paired supervision, we leverage CycleGAN to perform unpaired image-to-image translation, mapping abundant visible-light PCB images into the infrared domain. This generative process synthesizes high-fidelity pseudo-IR samples that preserve the structural semantics of defects while accurately simulating thermal distribution patterns. Subsequently, we construct a heterogeneous training strategy that fuses generated pseudo-IR data with limited real IR samples to train a lightweight YOLOv8 detector. Experimental results demonstrate that this method effectively enhances feature learning under low-data conditions. The augmented detector significantly outperforms models trained on limited real data alone and approaches the performance benchmarks of fully supervised training, proving the efficacy of pseudo-IR synthesis as a robust augmentation strategy for industrial inspection.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This applies CycleGAN for unpaired visible-to-IR translation plus YOLOv8 to boost PCB defect detection with scarce real IR data, but the abstract gives no metrics or validation to support the claims.

read the letter

The main thing here is a practical pipeline that uses CycleGAN to turn visible PCB photos into synthetic infrared images without paired training data, then mixes those with limited real IR samples to train a YOLOv8 detector for defects. The goal is to ease the data scarcity problem in industrial PCB inspection. That setup is the core contribution. The paper does a solid job naming a real bottleneck in manufacturing quality control, where collecting enough thermal images is expensive and slow. Framing the solution around unpaired translation and a mixed training strategy is a reasonable, off-the-shelf way to get more training volume without new hardware. It stays grounded by sticking to established models rather than claiming a new framework. The heterogeneous fusion step shows some thought about how to blend the generated and real data effectively. The work is not novel in its components. CycleGAN for image translation and YOLO for detection are both standard, so this is an application study rather than a technical advance. That is fine for a domain-specific paper, but it means the value rests entirely on whether the experiments hold up. The soft spots are clear and fairly large. The abstract asserts that the augmented detector outperforms limited-real-data baselines and approaches full supervision, yet it provides zero numbers, no dataset sizes, no ablation results, and no checks on whether the generated images actually keep defect locations and types intact. The central assumption that CycleGAN preserves structural semantics while simulating realistic thermal patterns is left untested in the text. Without metrics like SSIM on held-out pairs, FID scores, or expert agreement on defect labels in the pseudo samples, it is impossible to tell if gains come from better data or just more of it. CycleGAN often adds artifacts on small localized features, so this gap matters. This paper is for engineers and applied researchers working on industrial defect detection who need augmentation ideas for multimodal data. A reader facing similar low-data IR problems might pick up the overall strategy as a starting point. I would recommend sending it for peer review. The application is relevant enough that referees can evaluate the experiments once the missing quantitative details and validation steps are added. The idea is straightforward and worth that feedback.

Referee Report

3 major / 1 minor

Summary. The paper proposes a cross-modal data augmentation framework that uses CycleGAN for unpaired translation of visible-light PCB images into pseudo-infrared samples, which are then combined with limited real IR data to train a YOLOv8 detector for defect detection. It claims that this approach enhances feature learning under low-data conditions, significantly outperforms models trained only on limited real IR data, and approaches the performance of fully supervised training.

Significance. If the experimental claims hold, the work would offer a practical solution to IR data scarcity in industrial PCB inspection by leveraging abundant visible images via unpaired translation. This could reduce reliance on costly paired IR datasets and improve detector robustness, with potential applicability to other thermal imaging domains where semantic preservation during translation is feasible.

major comments (3)

[Abstract] Abstract: The abstract asserts that the augmented detector 'significantly outperforms models trained on limited real data alone' and 'approaches the performance benchmarks of fully supervised training,' yet provides no quantitative metrics (e.g., mAP, precision, recall), dataset sizes, ablation studies, or experimental protocols to support these claims.
[Method] Method/Experiments: The central assumption that CycleGAN produces high-fidelity pseudo-IR samples preserving exact defect locations and types requires validation, but the manuscript supplies no quantitative fidelity checks such as FID scores, SSIM, cycle-consistency loss on held-out data, or expert agreement on defect labels in generated images.
[Experiments] Experiments: No ablation isolating the contribution of generated pseudo-IR data quality versus simple data-volume increase is reported, making it impossible to confirm that performance gains stem from semantic preservation rather than generic augmentation effects.

minor comments (1)

[Method] The heterogeneous training strategy description lacks details on the fusion ratio between pseudo-IR and real IR samples or any regularization to mitigate domain shift.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will revise the paper to strengthen the presentation of results and validation.

read point-by-point responses

Referee: [Abstract] Abstract: The abstract asserts that the augmented detector 'significantly outperforms models trained on limited real data alone' and 'approaches the performance benchmarks of fully supervised training,' yet provides no quantitative metrics (e.g., mAP, precision, recall), dataset sizes, ablation studies, or experimental protocols to support these claims.

Authors: We agree that the abstract should include quantitative support for the claims. In the revised manuscript we will add specific metrics (mAP@0.5, precision, recall) for the proposed method versus the limited-real-IR baseline and the fully supervised upper bound, together with the exact counts of real IR images and generated pseudo-IR samples used in each training regime. revision: yes
Referee: [Method] Method/Experiments: The central assumption that CycleGAN produces high-fidelity pseudo-IR samples preserving exact defect locations and types requires validation, but the manuscript supplies no quantitative fidelity checks such as FID scores, SSIM, cycle-consistency loss on held-out data, or expert agreement on defect labels in generated images.

Authors: We accept that quantitative fidelity validation is needed. We will add FID and SSIM scores computed between generated pseudo-IR images and a held-out set of real IR images, report the cycle-consistency loss on that set, and include a brief qualitative analysis confirming that defect locations and types remain consistent after translation. revision: yes
Referee: [Experiments] Experiments: No ablation isolating the contribution of generated pseudo-IR data quality versus simple data-volume increase is reported, making it impossible to confirm that performance gains stem from semantic preservation rather than generic augmentation effects.

Authors: We agree that an ablation isolating the effect of semantic preservation is required. We will add an ablation study comparing three settings on the same limited real IR base: (i) limited real IR only, (ii) limited real IR plus an equal number of images augmented with standard geometric transforms, and (iii) limited real IR plus the CycleGAN-generated pseudo-IR samples. The results will be reported in a new table. revision: yes

Circularity Check

0 steps flagged

No circularity: standard external models applied without self-referential reduction

full rationale

The paper applies off-the-shelf CycleGAN for unpaired visible-to-IR translation and YOLOv8 for detection. No equations define outputs in terms of fitted inputs, no predictions reduce to training subsets by construction, and no load-bearing self-citations or uniqueness theorems are invoked. The augmentation strategy is presented as an empirical engineering choice whose efficacy is tested against external benchmarks (limited real data vs. full supervision), with no renaming of known results or ansatz smuggling. The derivation chain is self-contained against standard external components.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the unverified assumption that CycleGAN produces usable pseudo-IR data for this domain; no free parameters or new entities are introduced in the abstract.

axioms (1)

domain assumption Unpaired CycleGAN translation can map visible PCB images to infrared while preserving defect semantics and simulating thermal patterns
Invoked as the basis for the cross-modal data augmentation strategy in the abstract.

pith-pipeline@v0.9.0 · 5459 in / 1266 out tokens · 97565 ms · 2026-05-16T18:27:13.179965+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

[1]

Deep learning-based pcb defect detec- tion,

P. H. Huang, “Deep learning-based pcb defect detec- tion,” PhD thesis, University of Electronic Science and Technology of China, 2024

work page 2024
[2]

Study on ac- tive infrared thermal imaging method for solder ball defect detection,

X. N. Lu, Z. Z. He, N. N. Huet al., “Study on ac- tive infrared thermal imaging method for solder ball defect detection,”Transactions of the Chinese Society for Mechanical Engineering, vol. 52, no. 10, pp. 17– 24, 2016

work page 2016
[3]

Active thermal excitation-based infrared thermal imaging for deep detection of pipeline defects,

Z. Wang, Y . Zhang, Y . Yuet al., “Active thermal excitation-based infrared thermal imaging for deep detection of pipeline defects,”Acta Optica Sinica, vol. 38, no. 09, pp. 201–208, 2018

work page 2018
[4]

Research on pore de- fect identification in metallic materials based on in- 7 frared thermal imaging,

X. Chen, C. Yin, M. Xuet al., “Research on pore de- fect identification in metallic materials based on in- 7 frared thermal imaging,”Heat Treatment Technology, vol. 49, no. 12, pp. 66–70, 2020

work page 2020
[5]

Unpaired image-to-image translation using cycle-consistent ad- versarial networks,

J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent ad- versarial networks,” inIEEE International Conference on Computer Vision (ICCV), 2017, pp. 2223–2232

work page 2017
[6]

Unpaired image-to-image translation based on an improved cyclegan model,

J. He, F. Long, and X. Zhu, “Unpaired image-to-image translation based on an improved cyclegan model,” Journal of Yulin Normal University, vol. 39, no. 02, pp. 122–126, 2018

work page 2018
[7]

Image style transfer based on an improved cyclegan,

Z. Du, H. Shen, G. Song, and X. Li, “Image style transfer based on an improved cyclegan,”Optical Pre- cision Engineering, vol. 27, no. 08, pp. 1836–1844, 2019

work page 2019
[8]

Asymmetric cyclegan for unpaired nir-to-rgb face image transla- tion,

H. Dou, C. Chen, X. Hu, and S. Peng, “Asymmetric cyclegan for unpaired nir-to-rgb face image transla- tion,” inICASSP 2019 - 2019 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP), 2019, pp. 1757–1761

work page 2019
[9]

Implementation and research on mutual con- version between face sketches and photographs based on cyclegan,

F. Luo, “Implementation and research on mutual con- version between face sketches and photographs based on cyclegan,” PhD thesis, Chengdu University of Technology, 2021

work page 2021
[10]

Cyclegan-based visible-infrared image enhancement method for infrared power equipment object detec- tion,

J. Yin, Z. Li, L. Cui, W. Zhang, Q. Wang, and G. Si, “Cyclegan-based visible-infrared image enhancement method for infrared power equipment object detec- tion,” in2023 IEEE 5th International Conference on Power, Intelligent Computing and Systems (ICPICS), 2023, pp. 411–416

work page 2023
[11]

Research on cyclegan- based style transfer technology applied to face mask detection in facial images,

H. Guo, Z. Xiao, and P. Ji, “Research on cyclegan- based style transfer technology applied to face mask detection in facial images,”Journal of Qilu University of Technology, vol. 37, no. 01, pp. 15–20, 2023

work page 2023
[12]

Fuzzy-cycle: Visible to infrared ship image translation based on cyclegan,

L. Heet al., “Fuzzy-cycle: Visible to infrared ship image translation based on cyclegan,” in2024 6th In- ternational Conference on Industrial Artificial Intelli- gence (IAI), 2024, pp. 1–6. 8

work page 2024

[1] [1]

Deep learning-based pcb defect detec- tion,

P. H. Huang, “Deep learning-based pcb defect detec- tion,” PhD thesis, University of Electronic Science and Technology of China, 2024

work page 2024

[2] [2]

Study on ac- tive infrared thermal imaging method for solder ball defect detection,

X. N. Lu, Z. Z. He, N. N. Huet al., “Study on ac- tive infrared thermal imaging method for solder ball defect detection,”Transactions of the Chinese Society for Mechanical Engineering, vol. 52, no. 10, pp. 17– 24, 2016

work page 2016

[3] [3]

Active thermal excitation-based infrared thermal imaging for deep detection of pipeline defects,

Z. Wang, Y . Zhang, Y . Yuet al., “Active thermal excitation-based infrared thermal imaging for deep detection of pipeline defects,”Acta Optica Sinica, vol. 38, no. 09, pp. 201–208, 2018

work page 2018

[4] [4]

Research on pore de- fect identification in metallic materials based on in- 7 frared thermal imaging,

X. Chen, C. Yin, M. Xuet al., “Research on pore de- fect identification in metallic materials based on in- 7 frared thermal imaging,”Heat Treatment Technology, vol. 49, no. 12, pp. 66–70, 2020

work page 2020

[5] [5]

Unpaired image-to-image translation using cycle-consistent ad- versarial networks,

J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent ad- versarial networks,” inIEEE International Conference on Computer Vision (ICCV), 2017, pp. 2223–2232

work page 2017

[6] [6]

Unpaired image-to-image translation based on an improved cyclegan model,

J. He, F. Long, and X. Zhu, “Unpaired image-to-image translation based on an improved cyclegan model,” Journal of Yulin Normal University, vol. 39, no. 02, pp. 122–126, 2018

work page 2018

[7] [7]

Image style transfer based on an improved cyclegan,

Z. Du, H. Shen, G. Song, and X. Li, “Image style transfer based on an improved cyclegan,”Optical Pre- cision Engineering, vol. 27, no. 08, pp. 1836–1844, 2019

work page 2019

[8] [8]

Asymmetric cyclegan for unpaired nir-to-rgb face image transla- tion,

H. Dou, C. Chen, X. Hu, and S. Peng, “Asymmetric cyclegan for unpaired nir-to-rgb face image transla- tion,” inICASSP 2019 - 2019 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP), 2019, pp. 1757–1761

work page 2019

[9] [9]

Implementation and research on mutual con- version between face sketches and photographs based on cyclegan,

F. Luo, “Implementation and research on mutual con- version between face sketches and photographs based on cyclegan,” PhD thesis, Chengdu University of Technology, 2021

work page 2021

[10] [10]

Cyclegan-based visible-infrared image enhancement method for infrared power equipment object detec- tion,

J. Yin, Z. Li, L. Cui, W. Zhang, Q. Wang, and G. Si, “Cyclegan-based visible-infrared image enhancement method for infrared power equipment object detec- tion,” in2023 IEEE 5th International Conference on Power, Intelligent Computing and Systems (ICPICS), 2023, pp. 411–416

work page 2023

[11] [11]

Research on cyclegan- based style transfer technology applied to face mask detection in facial images,

H. Guo, Z. Xiao, and P. Ji, “Research on cyclegan- based style transfer technology applied to face mask detection in facial images,”Journal of Qilu University of Technology, vol. 37, no. 01, pp. 15–20, 2023

work page 2023

[12] [12]

Fuzzy-cycle: Visible to infrared ship image translation based on cyclegan,

L. Heet al., “Fuzzy-cycle: Visible to infrared ship image translation based on cyclegan,” in2024 6th In- ternational Conference on Industrial Artificial Intelli- gence (IAI), 2024, pp. 1–6. 8

work page 2024