Hsin-Lung Wu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Hsin-Lung Wu has a total of 24 co-authors. Here are the top co-authors by number of shared publications:
Jen-Chun Chang
National Taipei University
4 shared publications- Shi-Chun Tsai
National Chiao Tung University
3 shared publications Ching Chen
2 shared publicationsChun-Liang Jhong
Department of Computer Science and Information Engineering, National Taipei University, New Taipei City, Taiwan
2 shared publications- Yan-Jen Huang
Department of Computer Science and Information Engineering, National Taipei University, New Taipei City, Taiwan
2 shared publications Yun-Hong Chou
National Taipei University
2 shared publications
Latest Publications
- 26 June 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 331
Improving multimodal neural machine translation based on sparse attention
- 26 June 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 331
Improving multimodal neural machine translation based on sparse attention
- 29 January 2026
- Published by Elsevier in Neurocomputing
- Vol. 673
IPMT + + : Improving few-shot semantic segmentation with contrastive learning
- 29 January 2026
- Published by Elsevier in Neurocomputing
- Vol. 673
IPMT + + : Improving few-shot semantic segmentation with contrastive learning
- 5 December 2025
- Published by Elsevier in Signal Processing
- Vol. 242
Reversible data hiding for color images based on a frequency-first partial assignment strategy
- 5 December 2025
- Published by Elsevier in Signal Processing
- Vol. 242
Reversible data hiding for color images based on a frequency-first partial assignment strategy
- 7 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Semiconductor Manufacturing
- Vol. 39 (1) , 16-27
Artificial Intelligence–Based Evaluation System With Domain Adaptation for Ultrathin Wafer Detaping
As high-density packaging and low power consumption are prioritized in semiconductor production, ultrathin wafer processing has become vital. However, the fragile nature of these wafers often leads to damage during manual detaping, a process that is time-consuming, labor-intensive, and subjective. To address this, we developed an automated evaluation system integrating You Only Look Once version 8 (YOLOv8) with domain adaptation for robust performance under environmental variations. Through real-time image analysis, the system tracks detaping motion and quantifies operational stability using a finite state machine to calculate operation time. Experimental results under supervised learning showed average motion detection recall and precision of 99.06% and 98.33%, with a mean absolute error (MAE) of 0.17 s. Crucially, the proposed semi-supervised domain adaptive framework mitigated data distribution differences; in one experiment, mAP50 improved by 72.62% from baseline. For motion detection across two domain adaptation experiments, the system maintained high performance, with average precision and recall of 95.4% and 93.5% and an MAE of 0.275 s. This system provides an effective, automated evaluation of manual detaping, demonstrating the value of domain adaptation-based object detection in industrial applications.
- 7 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Semiconductor Manufacturing
- Vol. 39 (1) , 16-27
Artificial Intelligence–Based Evaluation System With Domain Adaptation for Ultrathin Wafer Detaping
As high-density packaging and low power consumption are prioritized in semiconductor production, ultrathin wafer processing has become vital. However, the fragile nature of these wafers often leads to damage during manual detaping, a process that is time-consuming, labor-intensive, and subjective. To address this, we developed an automated evaluation system integrating You Only Look Once version 8 (YOLOv8) with domain adaptation for robust performance under environmental variations. Through real-time image analysis, the system tracks detaping motion and quantifies operational stability using a finite state machine to calculate operation time. Experimental results under supervised learning showed average motion detection recall and precision of 99.06% and 98.33%, with a mean absolute error (MAE) of 0.17 s. Crucially, the proposed semi-supervised domain adaptive framework mitigated data distribution differences; in one experiment, mAP50 improved by 72.62% from baseline. For motion detection across two domain adaptation experiments, the system maintained high performance, with average precision and recall of 95.4% and 93.5% and an MAE of 0.275 s. This system provides an effective, automated evaluation of manual detaping, demonstrating the value of domain adaptation-based object detection in industrial applications.
- Open Access
- 25 January 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 21559-21568
Block-Wise Separable Convolutions: An Alternative Way to Factorize Standard Convolutions
In this paper, we introduce block-wise separable convolutions (BlkSConv) to replace the standard convolutions for compressing deep CNN models. First, BlkSConv expresses the standard convolutional kernel as an ordered set of block vectors each of which is a linear combination of fixed basis block vectors. Then it eliminates most basis block vectors and their corresponding coefficients to obtain an approximated convolutional kernel. Moreover, the proposed BlkSConv operation can be efficiently realized via a combination of pointwise and group-wise convolutions. Thus the constructed networks have smaller model size and fewer multiply-adds operations while keeping comparable prediction accuracy. We also develop a hyperparameter search framework based on principal component analysis (PCA) to determine a qualified hyperparameter setting of the block depth and number of basis block vectors. By this search framework, we construct networks which achieve nice prediction performance while simultaneously satisfying the constraints of model size and model efficiency. Our code, data, and models are available at https://github.com/yanjenhuang/blksconv.
- 25 January 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 21559-21568
Open AccessBlock-Wise Separable Convolutions: An Alternative Way to Factorize Standard Convolutions
In this paper, we introduce block-wise separable convolutions (BlkSConv) to replace the standard convolutions for compressing deep CNN models. First, BlkSConv expresses the standard convolutional kernel as an ordered set of block vectors each of which is a linear combination of fixed basis block vectors. Then it eliminates most basis block vectors and their corresponding coefficients to obtain an approximated convolutional kernel. Moreover, the proposed BlkSConv operation can be efficiently realized via a combination of pointwise and group-wise convolutions. Thus the constructed networks have smaller model size and fewer multiply-adds operations while keeping comparable prediction accuracy. We also develop a hyperparameter search framework based on principal component analysis (PCA) to determine a qualified hyperparameter setting of the block depth and number of basis block vectors. By this search framework, we construct networks which achieve nice prediction performance while simultaneously satisfying the constraints of model size and model efficiency. Our code, data, and models are available at https://github.com/yanjenhuang/blksconv.
- Open Access
- 16 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 130306-130318
New Coarse-to-Fine Approaches for Age Estimation Based on Separable Convolutions
In this paper, we study lightweight age estimation methods based on a coarse-to-fine approach in which the network performs age prediction with multiple stages. In each stage, the network only focuses on refining the coarse age prediction generated from the previous stage. The final age prediction is the combination of all staged prediction values. We observe that these stages have a causal relationship, that is, the output of each stage is highly correlated with outputs of its former stages. Thus, each stage should share the information of its previous stage before making a refined prediction. Based on this observation, we construct a new compact CNN model called Homologous Stagewise Regression Network (HSR-Net). In HSR-Net, each stage shares the information of the last convolutional layer and then generates its own refined value. In addition, HSR-Net also addresses the age group ambiguity problem by utilizing an easy dynamic range construction. In order to enhance the prediction performance of HSR-Nets, it is naive to increase the number of kernels in each convolutional layer of HSR-Nets. However, the constructed HSR-Net has extremely large parameter size. To address this problem, we propose the separable HSR-Nets (SepHSR-Nets) where standard convolutions are replaced by depth-wise separable convolutions in the convolutional layers of HSR-Nets. In general, the parameter size of SepHSR-Nets ranges from 10K to 75K without sacrificing prediction performance. Experimental results show that SepHSR-Nets achieve competitive performance compared with the state-of-the-art compact models. Our code, data, and models are available at https://github.com/yanjenhuang/hsr-net.
- 16 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 130306-130318
Open AccessNew Coarse-to-Fine Approaches for Age Estimation Based on Separable Convolutions
In this paper, we study lightweight age estimation methods based on a coarse-to-fine approach in which the network performs age prediction with multiple stages. In each stage, the network only focuses on refining the coarse age prediction generated from the previous stage. The final age prediction is the combination of all staged prediction values. We observe that these stages have a causal relationship, that is, the output of each stage is highly correlated with outputs of its former stages. Thus, each stage should share the information of its previous stage before making a refined prediction. Based on this observation, we construct a new compact CNN model called Homologous Stagewise Regression Network (HSR-Net). In HSR-Net, each stage shares the information of the last convolutional layer and then generates its own refined value. In addition, HSR-Net also addresses the age group ambiguity problem by utilizing an easy dynamic range construction. In order to enhance the prediction performance of HSR-Nets, it is naive to increase the number of kernels in each convolutional layer of HSR-Nets. However, the constructed HSR-Net has extremely large parameter size. To address this problem, we propose the separable HSR-Nets (SepHSR-Nets) where standard convolutions are replaced by depth-wise separable convolutions in the convolutional layers of HSR-Nets. In general, the parameter size of SepHSR-Nets ranges from 10K to 75K without sacrificing prediction performance. Experimental results show that SepHSR-Nets achieve competitive performance compared with the state-of-the-art compact models. Our code, data, and models are available at https://github.com/yanjenhuang/hsr-net.
- 1 April 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 32 (9) , 5888-5901
Grayscale-Invariant Reversible Data Hiding Based on Multiple Histograms Modification
Grayscale-invariant reversible data hiding (GI-RDH) in color images is a data embedding framework in which the grayscales of a marked color image must be identical to those of the host color image. Recently, some state-of-the-art GI-RDH schemes were proposed. However, their performance in embedding distortion is unsatisfactory. In order to obtain better image quality, a well-known histogram-shifting-based RDH method called multiple histograms modification (MHM) is considered. In this paper, we propose an MHM-based GI-RDH scheme. First, we modified our previous GI-RDH scheme using a multiple-histogram-shifting approach instead of a difference expansion approach. Next, we designed a procedure to select expansion–bin pairs for generated histograms to achieve low embedding distortion through further data embedding. Specifically, we analyzed the expected embedding distortion of our MHM-based GI-RDH scheme given any set of expansion–bin pairs. We then formulated an optimization problem called the GI-MHM minimization problem to identify the optimal expansion–bin pairs for further embedding tasks. Finally, we generated an approximated solution for the GI-MHM minimization problem and conducted the embedding task with these selected expansion–bin pairs. The experimental results revealed that the proposed GI-RDH scheme outperformed previous methods when the embedding capacity was small.
- 1 April 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 32 (9) , 5888-5901
Grayscale-Invariant Reversible Data Hiding Based on Multiple Histograms Modification
Grayscale-invariant reversible data hiding (GI-RDH) in color images is a data embedding framework in which the grayscales of a marked color image must be identical to those of the host color image. Recently, some state-of-the-art GI-RDH schemes were proposed. However, their performance in embedding distortion is unsatisfactory. In order to obtain better image quality, a well-known histogram-shifting-based RDH method called multiple histograms modification (MHM) is considered. In this paper, we propose an MHM-based GI-RDH scheme. First, we modified our previous GI-RDH scheme using a multiple-histogram-shifting approach instead of a difference expansion approach. Next, we designed a procedure to select expansion–bin pairs for generated histograms to achieve low embedding distortion through further data embedding. Specifically, we analyzed the expected embedding distortion of our MHM-based GI-RDH scheme given any set of expansion–bin pairs. We then formulated an optimization problem called the GI-MHM minimization problem to identify the optimal expansion–bin pairs for further embedding tasks. Finally, we generated an approximated solution for the GI-MHM minimization problem and conducted the embedding task with these selected expansion–bin pairs. The experimental results revealed that the proposed GI-RDH scheme outperformed previous methods when the embedding capacity was small.
- Open Access
- 4 October 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 9, 137005-137014
Exploring More Capacity for Grayscale-Invariance Reversible Data Hiding
Grayscale-invariant reversible data hiding (RDH) in color images was developed recently in which the grayscales of the marked color image are required to be the same as that of the host color image. Recently, some grayscale-invariant RDH schemes were proposed but their performance on embedding capacity and distortion is not satisfactory due to the large number of correction bits generated for losslessly recovering the host green scales. In this work, we propose a grayscale-invariant RDH method which reduces the need for correction bits. We design a self-correcting mechanism which modifies the green scales for keeping grayscales unchanged while remembering the information of the host green scales. To obtain this task, we observe that all previous methods generated correction bits in a functional approach which results in many correction bits. Instead of the functional approach, we use a relational approach to preserve the grayscales. By this approach, the proposed algorithm only has to generate a small number of correction bits for recovering green scales. Moreover, this relational approach allows the algorithm to embed more message bits into some green scales. The experimental results show that the proposed method significantly improves the embedding capacity and the rate-distortion performance compared to previous works.
- 4 October 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 9, 137005-137014
Open AccessExploring More Capacity for Grayscale-Invariance Reversible Data Hiding
Grayscale-invariant reversible data hiding (RDH) in color images was developed recently in which the grayscales of the marked color image are required to be the same as that of the host color image. Recently, some grayscale-invariant RDH schemes were proposed but their performance on embedding capacity and distortion is not satisfactory due to the large number of correction bits generated for losslessly recovering the host green scales. In this work, we propose a grayscale-invariant RDH method which reduces the need for correction bits. We design a self-correcting mechanism which modifies the green scales for keeping grayscales unchanged while remembering the information of the host green scales. To obtain this task, we observe that all previous methods generated correction bits in a functional approach which results in many correction bits. Instead of the functional approach, we use a relational approach to preserve the grayscales. By this approach, the proposed algorithm only has to generate a small number of correction bits for recovering green scales. Moreover, this relational approach allows the algorithm to embed more message bits into some green scales. The experimental results show that the proposed method significantly improves the embedding capacity and the rate-distortion performance compared to previous works.
- 19 January 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Theory
- Vol. 67 (7) , 4929-4935
On the Communication Complexity of AND Functions
Log-rank conjecture is one of challenging problems in communication complexity. It is known to hold for some special function classes such as XOR functions $\text {f}(\text {x}\oplus \text {y})$ when the outer function f is monotone, symmetric, $\mathsf {AC}^{0}$ , low $\mathbb {F}_{2}$ -degree, linear threshold or read-k. However, for AND functions $\text {f}(\text {x}\wedge \text {y})$ , Log-rank conjecture is only known to hold when the outer function is a monotone function or a function with small alternating numbers. In this paper, we show that Log-rank conjecture holds for AND functions whose outer functions are read-once $\mathbb {F}_{2}$ -degree polynomials, constant- $\mathbb {F}_{2}$ -degree polynomials, and symmetric functions. For read-once and symmetric AND functions, their proof are based on the elementary analysis of Möbius sparsity of the outer functions and designing efficient protocols for computing the corresponding AND functions in terms of the Möbius sparsity of their outer functions. For a constant- $\mathbb {F}_{2}$ -degree AND function $\text {f}(\text {x}\wedge \text {y})$ , we use the concept of the variable rank to design a decision tree for computing f where the tree depth is bounded above by a polynomial in the logarithm of the Möbius sparsity of f. The communication protocol can be obtained easily from the constructed decision tree. Our results deepen the understanding of the log-rank conjecture for AND functions.
- 19 January 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Theory
- Vol. 67 (7) , 4929-4935
On the Communication Complexity of AND Functions
Log-rank conjecture is one of challenging problems in communication complexity. It is known to hold for some special function classes such as XOR functions $\text {f}(\text {x}\oplus \text {y})$ when the outer function f is monotone, symmetric, $\mathsf {AC}^{0}$ , low $\mathbb {F}_{2}$ -degree, linear threshold or read-k. However, for AND functions $\text {f}(\text {x}\wedge \text {y})$ , Log-rank conjecture is only known to hold when the outer function is a monotone function or a function with small alternating numbers. In this paper, we show that Log-rank conjecture holds for AND functions whose outer functions are read-once $\mathbb {F}_{2}$ -degree polynomials, constant- $\mathbb {F}_{2}$ -degree polynomials, and symmetric functions. For read-once and symmetric AND functions, their proof are based on the elementary analysis of Möbius sparsity of the outer functions and designing efficient protocols for computing the corresponding AND functions in terms of the Möbius sparsity of their outer functions. For a constant- $\mathbb {F}_{2}$ -degree AND function $\text {f}(\text {x}\wedge \text {y})$ , we use the concept of the variable rank to design a decision tree for computing f where the tree depth is bounded above by a polynomial in the logarithm of the Möbius sparsity of f. The communication protocol can be obtained easily from the constructed decision tree. Our results deepen the understanding of the log-rank conjecture for AND functions.
- 11 November 2018
- Published by Springer Nature
- p. 35-44
A New Approximation Algorithm for the d-dimensional Knapsack Problem Based on Hopfield Networks
- 11 November 2018
- Published by Springer Nature
- p. 35-44
A New Approximation Algorithm for the d-dimensional Knapsack Problem Based on Hopfield Networks