Jirayu Petchhan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Jirayu Petchhan has a total of 6 co-authors. Here are the top co-authors by number of shared publications:
Shun-Feng Su
2 shared publications- Shun-Feng Su
Department of Electrical Engineering, National Taiwan University of Science and Technology, Taipei, Taiwan
2 shared publications - Jenq-Haur Wang
Department of Computer Science and Information Engineering, National Taipei University of Technology, Taipei 106, Taiwan
1 shared publication - Surasachai Doungtap
International Graduate Program of Electrical Engineering and Computer Science, National Taipei University of Technology, Taipei 10608, Taiwan
1 shared publication - Surasachai Doungtap1 shared publication
- Varinya Phanichraksaphong
International Graduate Program of Electrical Engineering and Computer Science, National Taipei University of Technology, Taipei 10608, Taiwan
1 shared publication
Latest Publications
- Open Access
- 1 December 2025
- Published by Institute of Advanced Engineering and Science in Bulletin of Electrical Engineering and Informatics
- Vol. 14 (6) , 4713-4731
On-edge 2D-to-3D generative pipeline for seamless instance transformation
Despite ongoing challenges with fragmented workflows, latency in device imports, and the main issue of limitations in object reconstruction functionality, relying on imperfect extraction networks remains an impractical solution for scalable object generation. To deal with these constraints, we proposed an end-to-end pipeline that leverages a re-designed self-consistency mechanism—aimed at reducing discrimination, along with the beneficial enhancement from level-set projection and gradient-surface orthogonality. In addition, our approach designs dynamic 3D object creation with minimal manual effort by unifying surface topology and optimizing data loading, enabling a streamlined reconstruction process and more flexible object projection. Our method supports rapid, resource-efficient mesh reconstruction and consistently demonstrates performance improvements across multiple instance benchmarks, covering virtual projection tasks. Improvements in mesh topology reconstruction, as measured by the L1 Chamfer distance (CD) metric, are consistently higher, while the system also achieves significant transmission speedups—up to 56.5×—near-instant importing—along with lowering latency in practical rendering on virtual reality (VR) devices. This result highlights that refining mesh binding improves re-creation fidelity. Our approach to scalability leads to faster user engagement and allows automated deployment without requiring human intervention during importing.
- 1 December 2025
- Published by Institute of Advanced Engineering and Science in Bulletin of Electrical Engineering and Informatics
- Vol. 14 (6) , 4713-4731
Open AccessOn-edge 2D-to-3D generative pipeline for seamless instance transformation
Despite ongoing challenges with fragmented workflows, latency in device imports, and the main issue of limitations in object reconstruction functionality, relying on imperfect extraction networks remains an impractical solution for scalable object generation. To deal with these constraints, we proposed an end-to-end pipeline that leverages a re-designed self-consistency mechanism—aimed at reducing discrimination, along with the beneficial enhancement from level-set projection and gradient-surface orthogonality. In addition, our approach designs dynamic 3D object creation with minimal manual effort by unifying surface topology and optimizing data loading, enabling a streamlined reconstruction process and more flexible object projection. Our method supports rapid, resource-efficient mesh reconstruction and consistently demonstrates performance improvements across multiple instance benchmarks, covering virtual projection tasks. Improvements in mesh topology reconstruction, as measured by the L1 Chamfer distance (CD) metric, are consistently higher, while the system also achieves significant transmission speedups—up to 56.5×—near-instant importing—along with lowering latency in practical rendering on virtual reality (VR) devices. This result highlights that refining mesh binding improves re-creation fidelity. Our approach to scalability leads to faster user engagement and allows automated deployment without requiring human intervention during importing.
- Open Access
- 25 July 2023
- Published by MDPI AG in Applied Sciences
- Vol. 13 (15) , 8571
Towards Digital Twins of 3D Reconstructed Apparel Models with an End-to-End Mobile Visualization
Digital twin technologies are still developing and are being increasingly leveraged to facilitate daily life activities. This study presents a novel approach for leveraging the capability of mobile devices for photo collection, cloud processing, and deep learning-based 3D generation, with seamless display in virtual reality (VR) wearables. The purpose of our study is to provide a system that makes use of cloud computing resources to offload the resource-intensive activities of 3D reconstruction and deep-learning-based scene interpretation. We establish an end-to-end pipeline from 2D to 3D reconstruction, which automatically builds accurate 3D models from collected photographs using sophisticated deep-learning techniques. These models are then converted to a VR-compatible format, allowing for immersive and interactive experiences on wearable devices. Our findings attest to the completion of 3D entities regenerated by the CAP–UDF model using ShapeNetCars and Deep Fashion 3D datasets with a discrepancy in L2 Chamfer distance of only 0.089 and 0.129, respectively. Furthermore, the demonstration of the end-to-end process from 2D capture to 3D visualization on VR occurs continuously.
- 25 July 2023
- Published by MDPI AG in Applied Sciences
- Vol. 13 (15) , 8571
Open AccessTowards Digital Twins of 3D Reconstructed Apparel Models with an End-to-End Mobile Visualization
Digital twin technologies are still developing and are being increasingly leveraged to facilitate daily life activities. This study presents a novel approach for leveraging the capability of mobile devices for photo collection, cloud processing, and deep learning-based 3D generation, with seamless display in virtual reality (VR) wearables. The purpose of our study is to provide a system that makes use of cloud computing resources to offload the resource-intensive activities of 3D reconstruction and deep-learning-based scene interpretation. We establish an end-to-end pipeline from 2D to 3D reconstruction, which automatically builds accurate 3D models from collected photographs using sophisticated deep-learning techniques. These models are then converted to a VR-compatible format, allowing for immersive and interactive experiences on wearable devices. Our findings attest to the completion of 3D entities regenerated by the CAP–UDF model using ShapeNetCars and Deep Fashion 3D datasets with a discrepancy in L2 Chamfer distance of only 0.089 and 0.129, respectively. Furthermore, the demonstration of the end-to-end process from 2D capture to 3D visualization on VR occurs continuously.
- 16 February 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 69 (3) , 353-365
High-Intensified Resemblance and Statistic-Restructured Alignment in Few-Shot Domain Adaptation for Industrial-Specialized Employment
The industrial status quo has been changed hastily. Facilities require new technological factors and are less datadriven to make precise inferences at most. Thus, the use of virtual knowledge can be adapted to practical applications to limit human intervention in explanation, namely, comprehensive digital transformation. Deep transfer learning has the role of enabling transferable knowledge and awareness practice of annotating endpoints. However, practical transfer tasks are burdened while considering scarce or non-annotated instances accumulated from external sampling, and these may inevitably degenerate knowledge during training and further degrade performance. In this study, a novel framework composed of the spectral correlation alignment is proposed to enable spectral patterned structure for statistical criterion to diminish the domain disparity under a fewshot alignment and multiple attention orchestration to accelerate a few-shot adaptation. This technique also increases highconcentrated recognition in cross-domain similarity verification. The demonstrations are conducted under public visual adaptation bench-marks and realistic deployment. The empirical experiment illustrates that our approach is better in efficacy and adaptability under data-limited conditions. Besides, in similar realistic applications, it is evident that the proposed scheme is deployable and can be re-practicable to less time as well as data consumption, yet is better in performance.
- 16 February 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 69 (3) , 353-365
High-Intensified Resemblance and Statistic-Restructured Alignment in Few-Shot Domain Adaptation for Industrial-Specialized Employment
The industrial status quo has been changed hastily. Facilities require new technological factors and are less datadriven to make precise inferences at most. Thus, the use of virtual knowledge can be adapted to practical applications to limit human intervention in explanation, namely, comprehensive digital transformation. Deep transfer learning has the role of enabling transferable knowledge and awareness practice of annotating endpoints. However, practical transfer tasks are burdened while considering scarce or non-annotated instances accumulated from external sampling, and these may inevitably degenerate knowledge during training and further degrade performance. In this study, a novel framework composed of the spectral correlation alignment is proposed to enable spectral patterned structure for statistical criterion to diminish the domain disparity under a fewshot alignment and multiple attention orchestration to accelerate a few-shot adaptation. This technique also increases highconcentrated recognition in cross-domain similarity verification. The demonstrations are conducted under public visual adaptation bench-marks and realistic deployment. The empirical experiment illustrates that our approach is better in efficacy and adaptability under data-limited conditions. Besides, in similar realistic applications, it is evident that the proposed scheme is deployable and can be re-practicable to less time as well as data consumption, yet is better in performance.
- 6 July 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 121-122
Image Blending-assisted Few-Shot Cross-Domain Similarity Learning and Adaptation Tasks for Ambiguous Hazardous Incidents
Nowadays, machine learning technology is growing exponentially, our research integrates AI and digital twin and/or transformation implemented in a wide range of industries. The issues can be seen in the fields which is the use of knowledge from both the virtual and physical world adapted all the way through. Part of obvious issue seems similar to deep transfer learning, such a domain shift occurrence during two domains. Hence, our proposed framework was developed to learn domain-invariant representation through Kernel Higher-order Tensor Matching (KHoM) and emphasized by cross-domain similarity learning via SoftTriple. Results, where are evaluated on public dataset and new fatal circumstance data, have been investigated that our framework is able to diminish discrepant domains from transferable on higher-level feature domain-invariant lightly on a less exemplary adaptations, but be obtained tremendously by the backing of the recognizability and realizing of object homogeneity through learning the likeness.
- 6 July 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 121-122
Image Blending-assisted Few-Shot Cross-Domain Similarity Learning and Adaptation Tasks for Ambiguous Hazardous Incidents
Nowadays, machine learning technology is growing exponentially, our research integrates AI and digital twin and/or transformation implemented in a wide range of industries. The issues can be seen in the fields which is the use of knowledge from both the virtual and physical world adapted all the way through. Part of obvious issue seems similar to deep transfer learning, such a domain shift occurrence during two domains. Hence, our proposed framework was developed to learn domain-invariant representation through Kernel Higher-order Tensor Matching (KHoM) and emphasized by cross-domain similarity learning via SoftTriple. Results, where are evaluated on public dataset and new fatal circumstance data, have been investigated that our framework is able to diminish discrepant domains from transferable on higher-level feature domain-invariant lightly on a less exemplary adaptations, but be obtained tremendously by the backing of the recognizability and realizing of object homogeneity through learning the likeness.