Zhiyong WANG bio photo

Email

Github

Google Scholar

About me

Hi, I am WANG Zhiyong. I just passed my PhD defense on this Aug, under the supervision of Prof. SO and Prof. Zoubir.
[Highlight] I am looking for a Post doc position. Contact me if you are interested in my works!

Academic Background

  • Sep 2020 - Aug 2023: City University of Hong Kong (PhD)

  • Sep 2017 - Jun 2020: Xi’an Jiaotong University (MS)

  • Sep 2013 - Jun 2017: Zhengzhou University (BS)


Research Interests

  • Sparse recovery

  • Robust signal processing

  • Matrix/tensor analysis (completion and robustness)

  • Applications to nature images, hyperspectral imaging and videos


Selected Publications

  • Z.-Y. Wang, H. C. So, and Z. Liu. “Fast and robust rank-one matrix completion via maximum correntropy criterion and half-quadratic optimization,” Signal Process., 2022.
    In this paper, the correntropy (the Welsch function) is adopted to measure the error for rank-one matrix completion, and half-quadratic optimization is utilized to convert the Welsch function into a penalty penalized least squares problem. We test the effectiveness of our method using natural images.

  • Z.-Y. Wang, X. P. Li, H. C. So and A. M. Zoubir, “Adaptive rank-one matrix completion using sum of outer products,” IEEE Trans. Circuits Syst. Video Technol., 2023.
    The conventional rank-one model recovers the missing entries via seeking a rank-one basis matrix at each outer iteration, and the basis matrix will not be further updated once it has been computed. In this paper, we find the basis vectors of the underlying matrix according to the observed entries, and gradually increase the vector number until an appropriate rank estimate is reached. Compared with the conventional model, our scheme performs completion from the vector viewpoint and is able to generate continuously updated rank-one basis matrices. It is proved that the recovery error of the developed rank-one scheme is no bigger than that of the conventional model. Experimental results show the superiority of our method over the competing techniques.

  • Z.-Y. Wang, X. P. Li and H. C. So, “Robust matrix completion based on factorization and truncated-quadratic loss function,” IEEE Trans. Circuits Syst. Video Technol., 2023.
    When the Weslch function is used for robustness, it down-weighs all data including normal data, then the truncated-quadratic function is adopted for robustness. The latter only penalizes outlier-contaminated data, and it achieves robustness via truncating the magnitudes of outlier-corrupted observations. As the truncated-quadratic function is nonconvex and nonsmooth, half-quadratic optimization is adopted. Then, the additive and multiplicative forms of the function are derived, resulting in two effective robust matrix completion algorithms based on factorization, namely, HOAT and HOMT. Numerical results using hyperspectral images demonstrate that our algorithms are superior to the state-ofthe-art methods in terms of restoration accuracy and runtime.

  • Z.-Y. Wang, X. P. Li, H. C. So and Z. Liu, “Robust PCA via non-convex half-quadratic regularization,” Signal Process., 2023.
    In this paper, a new nonconvex penalty called half-quadratic function is proposed, which is employed for robustness and sparsity. We derive the proximity operator of the developed penalty and apply it for robust principal component analysis. Experimental results based on real-world videos and face images demonstrate that the devised algorithm can effectively extract the low-rank and sparse components.

  • X. P. Li, Z.-Y. Wang, Z. L. Shi, H. C. So, N. D. Sidiropoulos, “Robust tensor completion via capped Frobenius norm,” IEEE Trans. Neural Netw. Learn. Syst., 2023.
    In this paper, a capped Frobenius norm is proposed, which is similar to the truncated-quadratic function, and half-quadratic function is exploited. We apply the developed norm to robust tensor completion, and experimental results based on color images and videos show the effectiveness of our method.

  • Z.-Y. Wang, H. C. So and A. M. Zoubir, “Robust low-rank matrix recovery via hybrid ordinary-Welsch function,” IEEE Trans. Signal Process., 2023.
    Since the truncated-quadratic function is nonsmooth and the Welsch function down-weighs all data, we devise a new loss function called hybrid ordinary-Welsch (HOW), which is smooth and only penalize the outlier-contaminated data. As HOW is nonconvex, the Legendre-Fenchel transform is exploited, which converts the problem as a sum of convex subproblems with closed-form solution. We apply HOW to factorization based matrix completion for robustness, and scaled alternating steepest descent is used to solve the corresponding factorization factors. Furthermore, since the implicit regularizer generated by HOW is sparsity-inducing regularizer, we also apply it to robust principal component analysis. Theoretical analyses associated with our algorithms are provided, and experimental results using hypersepctral images, videos and human faces show the effectiveness of our algorithm.


Patents


Selected Awards

  • Aug 2023：Outstanding Academic Performance Award (OAPA), City University of Hong Kong.

  • Aug 2022：Research Tuition Scholarship (RTS), City University of Hong Kong.

  • Jun 2022：First Prize in Research Student Symposium.