Graph Signal Wiener Filtering in the Linear Canonical Domain: Theory and Method Design
Cheng, Zhang
The graph linear canonical transform (GLCT)-based filtering methods often optimize transform parameters and filters separately, which results in high computational costs and limited stability. To address this issue, this paper proposes a trainable joint optimization framework that combines GLCT parameters and Wiener filtering into an end-to-end learning process, allowing for synergistic optimization between transform domain construction and filtering operations. The proposed method not only eliminates the cumbersome grid search required by traditional strategies but also significantly enhances the flexibility and training stability of the filtering system. Experimental results on real-world graph data show the proposed method outperforms existing methods in denoising tasks, featuring superior denoising performance, higher robustness and lower computational complexity.
academic
Graph Signal Wiener Filtering in the Linear Canonical Domain: Theory and Method Design
输入: 图信号f,目标信号f̃,参数网格A,B,D
输出: 最优参数(a*,b*,d*),最优滤波器H*
1. 预计算谱基(特征分解)
2. for a ∈ A, b ∈ B, d ∈ D:
- 构造GLCT算子F^M和F^{M^{-1}}
- 求解Wiener-Hopf方程: h = T^{-1}q
- 评估损失MSE(H,a,b,d)
- 更新最优解