A Deep Generative Approach for Neural Augmentation of Heat Sink Surface Defects

Jeongwoo Lee; Sooyoung Lee

doi:10.57062/ijpem-st.2025.00115

Int. J. Precis. Eng. Manuf.-Smart Tech. > Volume 3(2); 2025 > Article

Lee and Lee: A Deep Generative Approach for Neural Augmentation of Heat Sink Surface Defects

Regular Paper

International Journal of Precision Engineering and Manufacturing-Smart Technology 2025;3(2):151-159.

Published online: July 1, 2025

DOI: https://doi.org/10.57062/ijpem-st.2025.00115

A Deep Generative Approach for Neural Augmentation of Heat Sink Surface Defects

Jeongwoo Lee¹ · Sooyoung Lee^1.²

¹School of Mechanical Engineering, Chung-Ang University, 84 Heukseok-ro, Dongjak-gu, Seoul, 06974, Republic of Korea

²School of Computer Science and Engineering, Chung-Ang University, 84 Heukseok-ro, Dongjak-gu, Seoul, 06974, Republic of Korea

Sooyoung Lee, lees@cau.ac.kr

Received April 21, 2025 Revised June 13, 2025 Accepted June 16, 2025

This is an Open Access journal distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/4.0/) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

Abstract

The need to obtain a diverse set of high-fidelity defect cases has recently gained attention for enhancing defect detection performance. However, collecting actual defect samples remains challenging due to the high cost along with considerable data imbalance across different defect scenarios. To address these issues, this study proposed a deep learning-based approach for generating synthetic defect cases based on specified defect information. More specifically, a conditional generative adversarial network (cGAN) was developed to generate high-fidelity synthetic images of heat sink surface defects conditioned on user-defined defect types, locations, and sizes. The generative performance of the proposed model was quantitatively evaluated. The proposed model achieved a Learned Perceptual Image Patch Similarity (LPIPS) of 0.1738 and a Fréchet Inception Distance (FID) of 23.5749. Qualitative results validated that the proposed model successfully generated a variety of synthetic defect cases in accordance with the provided defect specifications. These findings demonstrate the potential of the proposed method as a neural augmentation strategy to enhance defect detection performance across various industrial applications.

Keywords: Deep learning · Generative adversarial network · Defect generation · Data augmentation

1 Introduction

Defects play a significant role in compromising product quality, often leading to degraded performance, reduced lifespan, and even functional failure [1–3]. Such defects, often characterized by surface and structural flaws—such as scratches, stains, cracks, and deformations—are considered critical concerns in various industries, including semiconductors, automotive, secondary batteries, and display technologies [4–8]. For example, surface defects in heat sinks are mainly caused by repeated thermal loading, accumulation of mechanical stresses, and manufacturing processes. This damage accumulates over time and progressively worsens under thermal fatigue conditions, ultimately degrading thermal performance [9–11]. Therefore, detection of these defects is essential for enhancing product reliability and optimizing process conditions.

Recent advances in deep learning have facilitated the development of diverse methods for surface defect inspection and anomaly detection. Several reconstruction-based approaches aim to model defect-free representations and identify anomalies by detecting deviations from the reconstructed outputs [12,13]. In addition, various supervised methods have been proposed, including a YOLO-based scheme for metal surface inspection [14], a convolutional neural network for joint detection and segmentation of surface cracks [15], an attention-enhanced multi-view fusion network for defect identification [16], and a ResNet-based classifier for multi-class defect categorization [17]. These aforementioned studies underscore the growing attention and rapid advancement of deep learning approaches specifically tailored for surface defect detection and visual anomaly identification across a broad range of application domains.

While data-driven approaches have been widely adopted to detect defects, obtaining sufficient and diverse defect cases remains a significant challenge [18,19]. Generally speaking, data-driven methods operate by identifying certain patterns from a large number of observations within a dataset. However, a fundamental issue in data-driven defect detection methods lies in the considerably smaller number of available defect cases compared to normal instances, leading to a severe data imbalance problem [20]. Such a data imbalance issue can inherently bias the data-driven model’s learning process toward characteristics of normal samples, due to the limited availability of defect cases, ultimately resulting in poor generalization performance for defect detection tasks [21]. To address this issue, the proposed generative approach aims to enrich the defect data distribution, thereby facilitating the training of downstream segmentation or classification models.

Conventional data augmentation methods typically rely on basic transformations such as rotation, brightness adjustment, and resizing. For instance, Taylor et al. [22] evaluated common data augmentation techniques, including image rotation, flipping, cropping, and color adjustments, showing that geometric transformations can enhance classification accuracy on coarse-grained image datasets. Work by Wang et al. [23] proposed a two-stage data augmentation strategy using only conventional transformations to augment a small set of crack samples, and expanded the original dataset into a substantially larger and more diverse training set. Even though several studies have shown promising results of data augmentation to some extent, it is important to note that data augmentation fundamentally aims to increase the diversity of training data by introducing cases that are not originally present in the dataset. However, conventional methods are often limited in their ability to generate novel and complex shapes or patterns beyond the existing data distribution [24,25].

Recently, deep learning (DL)-based approaches have been proposed as effective methods for augmenting defect cases. Unlike conventional techniques, DL-based methods leverage deeply structured neural networks to learn the underlying characteristics of defects, enabling the generation of more realistic and diverse synthetic samples. For instance, Niu et al. [26] proposed a generative adversarial network (GAN) model to generate high-quality and diverse defect images of the commutator cylinder surface in direct current (DC) motors, demonstrating effectiveness in generating diverse defect samples. Work by Jain et al. [27] also proposed a GAN-based approach to hot-rolled steel strips, showing that synthetic augmentation significantly improves classification performance. Even though the aforementioned studies have shown promising augmentation performance using DL-based methods, the existing approaches lack the generation capability to explicitly condition the spatial and structural information of defects, which motivated this study.

In this study, we propose a deep learning (DL)-based approach for neural augmentation, generating synthetic defect cases by conditioning on the spatial and structural characteristics of defects. Specifically, we introduce a conditional generative adversarial network (cGAN) to synthesize surface defects of a heat sink, focusing on various combinations of defect location, size, shape, and type. To assess the visual similarity and distributional alignment between the generated and actual defects, we quantitatively evaluate our proposed method using Learned Perceptual Image Patch Similarity (LPIPS) and Fréchet Inception Distance (FID) metrics. Furthermore, we qualitatively demonstrate the effectiveness of the proposed model, demonstrating its ability to generate a diverse range of realistic defect cases that are consistent with the specified defect attributes. The main contributions of this study can be summarized as follows:

1) A conditional GAN framework is presented in which the generator is driven by a three-channel, one-hot-encoded label map, enabling deterministic, class-controlled synthesis of defects with user-specified location, shape, and type.
2) A patch-based discriminator operating on sub-receptive fields is incorporated to enhance the model’s ability to preserve fine-grained details associated with different surface defect categories.

The remainder of this paper is organized as follows: Section 2 provides detailed descriptions of the utilized dataset as well as the proposed method, including model architecture and learning strategy. Section 3 presents quantitative and qualitative analyses of the proposed approach in terms of defect generation performance. Finally, Section 4 concludes the paper.

2 Materials and Methods

In this section, we briefly describe the dataset utilized for synthetic defect generation, and subsequently provide a comprehensive explanation of the proposed methodology.

2.1 Data Description

In this study, we focus on the surface defects observed on the gold-plated tungsten-copper alloy heat sink. The utilized dataset [7] was acquired with a DAHENG MER-2000-5GC industrial camera equipped with a telecentric lens and illuminated by a ring-type LED light source. The raw images were pre-processed by Otsu binarization, morphological closing, contour extraction, and 320 × 320 cropping, and finally pixel-wise annotated for the three defect classes. While the heat sink is an essential component for dissipating heat from electronic chips, it should be noted that surface defects of the heat sink can degrade thermal conductivity and thereby increase the risk of chip failure due to overheating [28]. Fig. 1 visualizes several representative examples of heat sink surface defects, which can be categorized into three types: scratch, stain, and scratch+stain. The scratch defect represents a physical abrasion on the metal surface caused by mechanical damage, which can disrupt the heat transfer pathway and directly degrade thermal management performance. Furthermore, the stain defect represents a surface non-uniformity case induced by specific localized contamination, which can lead to reduced thermal conductivity as well as increased risk of poor contact during packaging processes. Lastly, defect cases showing both scratch and stain can be regarded as a more severe form of surface failure.

Detailed information about the dataset used in this study is given as follows. The entire dataset consists of 1,000 pairs of defect samples (scratch, stain, and scratch+stain) and their corresponding pixel-wise annotations. Each image has a resolution of pixels, and the associated label maps contain three distinct classes: defect-free regions, scratch regions, and stain regions. For compatibility to reduce computational overhead, all images and their corresponding label maps were resized to 256 × 256 pixels prior to training. The label maps are preprocessed into three channels using a one-hot encoding scheme, where each channel represents background, scratch, and stain information. We employ the multi-channel representation of the label maps to guide the model’s learning process in capturing defect-specific spatial features, including the presence, type, and location of certain defects. While the dataset is normalized to the range from −1 to 1, all samples are stratified into 85 for training, 5 for validation, and 10 for testing, respectively.

2.2 Proposed Model for Surface Defect Generation

In this work, we propose a DL-based approach for generating synthetic defect cases given defect-specific spatial information. Specifically, we develop a cGAN model to generate realistic defect samples from structured defect annotations. The GAN framework is designed to approximate the model distribution to the target data distribution through adversarial training between two sub-networks [29,30], generally consisting of a generator network G and a discriminator network D. First, given three-channel label maps containing defect-related spatial information, the generator network G is designed to produce synthetic defect cases through a series of two-dimensional convolutional layers in the contracting path and several transposed convolutional layers in the expanding path. Herein, we employ skip connection mechanisms between the cont racting and expanding paths to preserve defect-related spatial information in the generation process. In contrast, the discriminator network D is trained to determine whether the fed sample is actual or generated. To achieve this, we construct the input to D by concatenating the generated or actual image with its corresponding three-channel label map along the channel dimension. We design the discriminator network D as a patch-based model to effectively capture localized defect patterns spread across the spatial region. This patch-level discrimination process is developed to enable the generator network G to learn more feasible and localized details of the defects.

Fig. 2 provides a schematic description of the proposed neural network architectures for the sub-networks G and D. Given the one-hot encoded three-channel label map as input, the generator network G is trained to progressively reduce spatial dimensions while capturing high-level semantic feature representations through successive six convolutional layers. In the expanding path, the generator network enlarges the extracted feature maps using transposed convolutional layers, while feature maps from the corresponding contracting path are combined through skip connections to preserve spatial information and improve training stability [31,32]. Unlike conventional cGANs that incorporate random noise for sample generation, our framework adopts a noise-free conditional GAN specifically tailored for image-to-image generation tasks. This design choice aligns with the objective of the present study, which is to accurately capture the spatial defect-related characteristics specified by the input label map. Subsequently, the actual data and the synthetic samples generated by the generator are fed into the discriminator network D, which produces patch-wise probability maps through a series of four convolutional layers.

2.3 Learning Scheme

In order to train the sub-networks effectively, we propose a multi-part loss function, which can be divided into three components. First, the adversarial loss L_adv is defined by binary cross entropy (BCE) to measure the similarity between actual and generated samples based on binary classification labels. L_adv can be expressed as:

(1)

Ladv(G,D)=Ex,y[logD(x,y)]+Ex[log(1-D(x,G(x)))],

where x is the multi-channel label map and denotes the actual ground truth defect image. G(x) represents the synthetic defect image produced by the generator network conditioned on the label map x. In addition, the pixel-level loss function L_pixel is defined to preserve the pixel-wise localized details of the generated defects:

(2)

Lpixel(G)=Ex,y[‖y-G(x)‖1]

To generate more fine-grained and visually consistent results, we incorporate a perceptual loss function L_perceptual, which is calculated by the hidden feature maps drawn from a pre-trained VGG19 network [33] as:

(3)

Lperceptual(G)=Ex,y[‖φ(y)-φ(G(x))‖1],

where φ(·) denotes the feature map extraction from the pre-trained VGG19 network. By leveraging abstract-level features from the internal layer of the pre-trained network, L_perceptual ensures that the generated results satisfy structural consistency while preserving realistic texture information of the defects. As a result, the total loss function L_total is defined as a weighted sum of the components:

(4)

Ltotal=Ladv+λpixelLpixel+λperceptualLperceptual

through the manual search process, we employ λ_pixel = 100, λ_perceptual = 10 to balance the effects of pixel matching and perceptual consistency.

Additional training settings for the proposed model are summarized as follows. The proposed networks are trained using the Adam Optimizer, with the learnable parameters updated via a mini-batch gradient descent algorithm using a batch size of 4. To increase learning stability, we adopt an asymmetric learning rate scheduling strategy, where the generator network is updated twice for every update of the discriminator network. Learning rates of the generator and discriminator are set to be 0.0002 and 0.0001, respectively. The model is implemented using Python 3.10.11/PyTorch 2.3.0 and trained on a NVIDIA GeForce RTX 4090 GPU.

3 Results and Discussion

In this section, we first describe the evaluation strategies, followed by both quantitative and qualitative analyses of the defect generation performance of the proposed method. In addition, we provide further validation results under extrapolated annotation cases and finally discuss potential strengths of our approach.

3.1 Evaluation Metrics

We evaluate the quality of the generated defects with quantitative measures as in the following. First, we consider a LPIPS, which is designed to compute the perceptual distance between two images in high-dimensional feature space [34]. While a lower LPIPS score indicates a higher visual similarity, we obtain the LPIPS value retrieved from the pre-trained convolutional neural network (CNN) model, i.e., AlexNet [35]. The formula can be expressed as:

(5)

LPIPS(G(x),y)=∑l1HlWl∑h,w‖wl⊙(φ^l(G(x))hw-φ^l(y)hw)‖22

where l is the layer index of CNN and (h,w) are the spatial coordinates of the layer. H_l and W_l denote the height and width of the feature map. Besides, φ̂ represents the normalized feature vector and is the learned layer-wise channel weights. Second, the FID [36] is utilized to measure the distance between the feature distributions of the actual and generated datasets. The FID score can be computed as:

(6)

FID2=‖m-mw‖22+Tr(C+Cw-2(CCW)12)

where m and m_w denote the mean vectors for the inception features of the actual and generated images, respectively. C and C_w are the covariance matrices of those features. In this wise, we evaluate the proposed model in terms of both the pairwise similarity between each actual and generated sample as well as the distributional similarity between the actual and generated datasets. While the model is trained in a patch-wise manner to preserve fine-grained details, both LPIPS and FID are computed at the whole-image level. These global metrics evaluate how accurately the generated image reflects the overall spatial distribution, shape, and defect class as specified in the input label map.

3.2 Defect Generation Performance

First, we quantitatively analyze perceptual and distributional similarity between the actual and generated defect cases. Fig. 3 overall provides a quantitative evaluation of the proposed model regarding two metrics. Figs. 3(a) and 3(b) visualize the LPIPS scores for the test and training sets, respectively, where the red dashed line in each plot represents the average LPIPS score for each dataset. It is observed that the proposed model achieves the average LPIPS scores of 0.1738 for the test set and 0.1187 for the training set, indicating a high perceptual similarity between the generated defect cases and the corresponding actual samples. Considering that the LPIPS score reflects feature-level similarity aligned with human visual perception, we find that the generated defect images for the test and training sets exhibit high perceptual fidelity to the corresponding actual defect cases. Additionally, we evaluate the FID performance of the proposed approach as the model’s learning progresses, to examine how closely the distribution of the generated cases aligns with that of the actual data. As shown in Fig. 3(c), one can find that the FID values gradually decrease as the training process converges, indicating that the distribution of generated images becomes increasingly similar to that of the actual data. We observe that the FID converges to 23.5749 for the test set and 5.4572 for the training set, showing a moderate gap between the two values without overfitting problem. To summarize, the results demonstrate that the generated defect cases not only exhibit perceptual consistency but also achieve distributional similarity compared to the actual defect cases.

Subsequently, we qualitatively assess the generation capability of the proposed model conditioned on different types of defects. First, Fig. 4 presents a qualitative comparison between the generated and actual defect cases, given the scratch label maps. While scratch defects—typically caused by physical contact or friction—tend to exhibit linear and sharp patterns, we find that the proposed model effectively captures their directionality, continuity, and contrast of the patterns, as appeared in the actual images. In particular, the enlarged region dominated by scratch defects shows similar spatial characteristics between the generated and actual samples.

Fig. 5 demonstrates the generation results of the proposed model in terms of stain defect cases. While stain defects are primarily caused by chemical residues or particles adhering to the surface, resulting in irregular and localized boundaries, we show that the proposed model is capable of generating such stain defect patterns with varying locations and sizes. It is noteworthy that the generated results closely resemble the actual data in terms of spatial distribution, both in cases where stain defects are scattered and where they are clustered in specific regions. This confirms the proposed model’s ability to effectively capture the diverse geometric characteristics of stain defects and to generate the appropriate defect samples according to different spatial configurations.

Lastly, Fig. 6 shows a qualitative evaluation of the generated results where both scratches and stains coexist. This example represents more generalized anomalies involving multiple defect types. As illustrated in the figure, the proposed model successfully captures the linear nature of the scratches as well as the spatial distribution of the stains. These results suggest that the model is able to reproduce the characteristics of each defect simultaneously, even under complex conditions involving multiple defects. Considering that such multiple defects can occur in real-world industrial environments, the proposed model’s capability to generate realistic multiple defect cases suggests its practicality. These characteristics further suggest the potential for future expansion into diverse industrial applications for various real-world defect augmentation tasks.

3.3 Validation for Extrapolated Annotation

Another key aspect of the defect generation method lies in its ability to generate desired defect patterns based on extrapolated conditions. This capability can enhance the controllability of the generation process, enabling the augmentation of arbitrary and user-specified defect cases. To validate this capability, the model’s generation performance is assessed under conditions where extrapolated human-annotated label maps are provided as input. The main purpose of this evaluation is to demonstrate how well the model reflects defect characteristics even with novel geometries or boundaries that are not included in the original dataset. As shown in Fig. 7, The generated defects successfully reflect the characteristics specified by the label maps, indicating that the model exhibits a high degree of controllability conditioned on the input conditions. We find that the generated defects are able to reproduce spatial characteristics even when scratches or stains cover significantly larger areas than those present in the original dataset. This suggests that the proposed model maintains high generalization performance across different scales and variations of defects, thereby offering the potential to reduce time and cost associated with defect data acquisition processes.

3.4 Potential Strength of the Proposed Approach

We further discuss the potential advantage of the proposed approach as a neural augmentation method. Results demonstrate that the proposed method is capable of generating defect images that accurately correspond to specific defect types, sizes, and spatial locations, enabling targeted data synthesis for diverse scenarios. In real-world applications, acquiring a diverse set of defect cases requires expensive, time-consuming, and labor-intensive procedures. Conventional data augmentation techniques, such as cropping, rotation, tilting, and brightness adjustment, mainly rely on basic spatial transformations and fail to adequately capture the intrinsic characteristics of defect patterns. On the other hand, the proposed approach is capable of synthesizing novel defect cases with varying locations, sizes, shapes, and types, thereby effectively reflecting the complex and irregular defect characteristics that may occur in real-world scenarios. Furthermore, the proposed neural network architecture is easily extensible to various image-based inspection domains, e.g., thermal imaging and biomedical imaging. Along with the ability to generate defects based on user-defined label maps, the proposed approach can directly contribute to improving the performance of data-driven defect detection models by addressing data imbalance through effective data augmentation. These potential strengths of the neural augmentation method suggest future development in a wide range of industrial disciplines.

4 Conclusions

In this study, we proposed a DL-based approach for neural augmentation of heat sink surface defects, enabling the generation of diverse and realistic synthetic defect cases under specified annotation conditions. Mainly, we developed a cGAN model to achieve effective generation of realistic defects under user-specified conditions. Both quantitative and qualitative results showed that our proposed method can generate high-fidelity data for the desired defect patterns. Specifically, our model achieved an LPIPS score of 0.1738 and FID score of 23.5749, indicating that the generated defect cases exhibit a high level of perceptual fidelity while achieving distributional similarity with respect to the actual defect cases. In addition, the model demonstrated robust performance under extrapolated annotation conditions, successfully generating diverse and spatially consistent heat sink defect cases based on user-specified arbitrary label maps. Furthermore, we presented the potential strength of the proposed approach as a neural augmentation method, which can significantly reduce the cost and time required to acquire defect samples and further address the data imbalance problem. Future research directions include further validation across a wider range of defect types and manufacturing environments to enhance the model’s generality and practical applicability.

Declarations

Acknowledgement(s)

This work was supported by the National Research Foundation of Korea (No. RS-2024-00342929) funded by the Korea Ministry of Science and ICT.

Fig. 1

Several examples of heat sink surface defects and their corresponding label maps in three different groups: scratch, stain, and scratch+stain cases

Fig. 2

Architectural description of the generator and discriminator networks for surface defect generation

Fig. 3

Quantitative evaluation of the proposed surface defect image generation model. (a) and (b) show the LPIPS scores for each sample in the test and train sets, respectively, along with their distributions as histograms. (c) presents the FID scores over epochs for both the test and train sets

Fig. 4

Qualitative comparison between the generated and actual defect cases from the test set (scratch). Note that the generated images are conditioned on the given label maps

Fig. 5

Qualitative comparison between the generated and actual defect cases from the test set (stain). Note that the generated images are conditioned on the given label maps

Fig. 6

Qualitative comparison between the generated and actual defect cases from the test set (scratch+stain). Note that the generated images are conditioned on the given label maps

Fig. 7

Several examples showing the generated defect cases under extrapolated annotation label maps

References

1. Ren, Z., Fang, F., Yan, N. & Wu, Y. (2022). State of the art in defect detection based on machine vision. International Journal of Precision Engineering and Manufacturing-Green Technology, 9(2), 661–691.

2. Lv, Z., Lu, Z., Xia, K., Zuo, H., Jia, X., Li, H. & Xu, Y. (2024). Laacnet: Lightweight adaptive activation convolution network-based defect detection on polished metal surface. Engineering Applications of Artificial Intelligence, 133, 108482.

3. Han, M., Zheng, H., Gao, Y., Zhou, Z. & Liu, X. (2024). Detection of near-surface defects in viscoelastic material based on focused ultrasound thermal effects. Ultrasonics, 143, 107415.

4. Zhou, Q., Chen, R., Huang, B., Liu, C., Yu, J. & Yu, X. (2019). An automatic surface defect inspection system for automobiles using machine vision methods. Sensors, 19(3), 644.

5. Ma, J., Zhang, T., Yang, C., Cao, Y., Xie, L., Tian, H. & Li, X. (2023). Review of wafer surface defect detection methods. Electronics, 12(8), 1787.

6. Tang, H., Liang, S., Yao, D. & Qiao, Y. (2023). A visual defect detection for optics lens based on the YOLOv5-C3CA-SPPF network model. Optics Express, 31(2), 2628–2643.

7. Yang, K., Liu, Y., Zhang, S. & Cao, J. (2022). Surface defect detection of heat sink based on lightweight fully convolutional network. IEEE Transactions on Instrumentation and Measurement, 71, 1–12.

8. Yousefian-Jazi, A., Ryu, J.-H., Yoon, S. & Liu, J. J. (2014). Decision support in machine vision system for monitoring of TFT-LCD glass substrates manufacturing. Journal of Process Control, 24(6), 1015–1023.

9. Sun, Y., Li, T., Lan, L., Chen, J., Zhu, W., Xue, S. & Jin, L. (2023). A mini-review on the thermal fatigue properties of copper materials applied at the front-end of synchrotron radiation facilities. Entropy, 25(5), 714.

10. Baig, M. M. A., Hassan, S. F., Saheb, N. & Patel, F. (2021). Metal matrix composite in heat sink application: Reinforcement, processing, and properties. Materials, 14(21), 6257.

11. Chen, J., & Yang, Z.-G. (2023). Cause analysis on abnormal failure of copper-substrate pin fin heat sink for new energy vehicle. Engineering Failure Analysis, 143, 106947.

12. Schlegl, T., Seeböck, P., Waldstein, S. M., Langs, G. & Schmidt-Erfurth, U. (2019). f-AnoGAN: Fast unsupervised anomaly detection with generative adversarial networks. Medical Image Analysis, 54, 30–44.

13. Zavrtanik, V., Kristan, M. & Skočaj, D. (2021). Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision; pp 8330–8339.

14. Lv, X., Duan, F., Jiang, J.-J., Fu, X. & Gan, L. (2020). Deep active learning for surface defect detection. Sensors, 20(6), 1650.

15. Tabernik, D., Šela, S., Skvarč, J. & Skočaj, D. (2020). Segmentation-based deep-learning approach for surface-defect detection. Journal of Intelligent Manufacturing, 31(3), 759–776.

16. Xie, L., Xiang, X., Xu, H., Wang, L., Lin, L. & Yin, G. (2020). FFCNN: A deep neural network for surface defect detection of magnetic tile. IEEE Transactions on Industrial Electronics, 68(4), 3506–3516.

17. Zhou, C., Wang, H., Liu, Y., Ni, X. & Liu, Y. (2022). Green plums surface defect detection based on deep learning methods. IEEE Access, 10, 100397–100407.

18. Shi, X., Zhang, S., Cheng, M., He, L., Tang, X. & Cui, Z. (2023). Few-shot semantic segmentation for industrial defect recognition. Computers in Industry, 148, 103901.

19. Huang, C.-H., Lee, P.-H., Chang, S.-H., Kuo, H.-C., Sun, C.-W., Lin, C.-C., Tsai, C.-L. & Liu, X. (2021). Automated optical inspection method for light-emitting diode defect detection using unsupervised generative adversarial neural network. Crystals, 11(9), 1048.

20. Chen, W., Yang, K., Yu, Z., Shi, Y. & Chen, C. P. (2024). A survey on imbalanced learning: latest research, applications and future directions. Artificial Intelligence Review, 57(6), 137.

21. He, X., Luo, Z., Li, Q., Chen, H. & Li, F. (2023). Dg-gan: A high quality defect image generation method for defect detection. Sensors, 23(13), 5922.

22. Taylor, L., & Nitschke, G. (2018). Improving deep learning with generic data augmentation. In: Proceedings of the 2018 IEEE Symposium Series on Computational Intelligence; pp 1542–1547.

23. Wang, Z., Yang, J., Jiang, H. & Fan, X. (2020). CNN training with twenty samples for crack detection via data augmentation. Sensors, 20(17), 4849.

24. Shorten, C., & Khoshgoftaar, T. M. (2019). A survey on image data augmentation for deep learning. Journal of Big Data, 6(1), 1–48.

25. Kumar, T., Brennan, R., Mileo, A. & Bendechache, M. (2024). Image data augmentation approaches: A comprehensive survey and future directions. IEEE Access, 12, 187536–187571.

26. Niu, S., Li, B., Wang, X. & Lin, H. (2020). Defect image sample generation with GAN for improving defect recognition. IEEE Transactions on Automation Science and Engineering, 17(3), 1611–1622.

27. Jain, S., Seth, G., Paruthi, A., Soni, U. & Kumar, G. (2022). Synthetic data augmentation for surface defect detection and classification using deep learning. Journal of Intelligent Manufacturing (pp. 1–14.

28. Gutiérrez-Posada, V., Akhavan-Safar, A., Simões, B., Carbas, R., Marques, E. & da Silva, L. (2025). Effect of bonding defects on heat transfer and creep response of microprocessor-heatsink adhesive joints. Journal of Materials Science: Materials in Electronics, 36(1), 1–13.

29. Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A. & Bengio, Y. (2014). Generative adversarial nets. Advances in Neural Information Processing Systems (pp. 27.

30. Creswell, A., White, T., Dumoulin, V., Arulkumaran, K., Sengupta, B. & Bharath, A. A. (2018). Generative adversarial networks: An overview. IEEE Signal Processing Magazine, 35(1), 53–65.

31. Ronneberger, O., Fischer, P. & Brox, T. (2015). U-net: Convolutional networks for biomedical image segmentation. In: Proceedings of the Medical Image Computing and Computer-Assisted Intervention 2015, part III 18, 9351.

32. He, K., Zhang, X.Ren, S. & Sun, J. (2016). Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; pp 770–778.

33. Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556.

34. Zhang, R., Isola, P.Efros, A. A.Shechtman, E. & Wang, O. (2018). The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; pp 586–595.

35. Krizhevsky, A., Sutskever, I. & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems (pp. 25.

36. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B. & Hochreiter, S. (2017). Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems (pp. 30.

Biography

Jeongwoo Lee is an undergraduate researcher in the School of Mechanical Engineering at Chung-Ang University, Seoul, Republic of Korea. He is currently affiliated with the Industrial AI Laboratory, where he focuses on the AI development to mechanical and manufacturing systems. His research interests include deep generative models and physics-informed machine learning for solving complex engineering problems.

Biography

Sooyoung Lee is currently an assistant professor and a principal investigator of the Industrial AI Laboratory at Chung-Ang University, Seoul, Republic of Korea. He received a B.S. degree from Chung-Ang University in 2019 and received his Ph.D. degree from Pohang University of Science and Technology (POSTECH) in 2023. He was an Honorary Research Associate for Sustainable Smart Manufacturing using AI with the University of Wisconsin-Madison, WI, USA, supported by the High-Potential Individuals Global Training Program of International Joint Research. He was also a Senior Research developing large-scale AI-based next-era semiconductor manufacturing technologies at Future Technology Institute, SK Hynix. His research interests span AI-enabled accelerated design, smart manufacturing and intelligent informatics for a variety of engineering systems and industrial applications.