News
New Publications of AVITECH Research Group at ICASSP 2026
Our research group continues to achieve notable results at the 2026 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP 2026), held in Barcelona, Spain, from May 4–8, 2026 — a Qualis A1 international conference and one of the world’s leading academic forums in signal processing, artificial intelligence, and machine learning, organized annually by the Institute of Electrical and Electronics Engineers Signal Processing Society. This year, the conference received more than 11,000 paper submissions from researchers worldwide, with an acceptance rate of approximately 40%, highlighting the highly competitive and prestigious nature of ICASSP.
Representing the group, Dr. Lê Trung Thành attended the conference and presented the group’s latest research works in tensor decomposition and multidimensional data processing. This year, the group had three papers accepted at ICASSP 2026, focusing on modern research directions in optimization, tensor learning, and large-scale data processing:
1. Dang, N.Q., Le, T.T., Trung, N.L., and Abed-Meraim, K., “Re-LL1: An Effective Regularized (L,L,1)-Tensor Decomposition Method for Video Background Modeling and Foreground Separation,” Proc. IEEE ICASSP 2026.
2. Dang, N.Q., Nhat, D.M., Le, T.T., Trung, N.L., and Abed-Meraim, K., “Fast and Robust Triple Tensor Decomposition With Data Corruption,” Proc. IEEE ICASSP 2026.
3. Lan, N.T.N., Le, T.T., Trung, N.L., and Abed-Meraim, K., “TriNet: A Novel and Memory-Efficient Tensor Network for Higher-order Tensor Decomposition,” Proc. IEEE ICASSP 2026.
These works aim to develop tensor models that are more efficient in both estimation accuracy and computational cost for practical applications such as surveillance video processing, data recovery, and large-scale multidimensional data analysis.
In particular, the Re-LL1 paper proposes a regularized tensor decomposition framework for background modeling and foreground separation in video sequences. The method integrates multiple regularization mechanisms to simultaneously exploit low-rank structure, sparsity, and temporal smoothness, thereby improving modeling capability and accelerating convergence.
The second paper introduces TriTD-ADMM, a robust tensor triple decomposition algorithm designed to handle corrupted data. The group developed the RPAS acceleration strategy, which significantly reduces computational complexity and enables the method to achieve superior processing speed across multiple benchmark datasets.
In the third research direction, the group proposed TriNet decomposition — a novel memory-efficient tensor network architecture for higher-order data. By employing relay factors to connect core tensors, TriNet reduces storage complexity while maintaining competitive estimation accuracy compared with state-of-the-art tensor methods.
The consecutive acceptance of multiple papers at ICASSP 2026 demonstrates the strong development of the research group in tensor methods, machine learning, and modern signal processing. It also highlights the group members’ research capability and growing international integration within the global scientific community.
THANH LE – DUC KIEN
