Since the introduction of finite rate of innovation (FRI) signals in 2002, various deterministic and stochastic techniques have been proposed to estimate the innovative parameters of FRI signal, e.g. time instants $t_k$ and weights $c_k$ belonging to a linear combination of finite number of Diracs, from its noisy samples regardless of sampling kernel's type. Having analyzed the Bayesian methods dedicated to the retrieval of signal innovations, particularly the IterML algorithm introduced by Wein & Srinivasan, we discover some limitations which still leave room for improvement. In this article, we present a novel stochastic hybrid algorithm utilizing both maximum–likelihood-estimation (MLE) and Modified Particle Swarm Optimization (MPSO) in order to improve upon IterML in terms of robustness to noise and accuracy of estimated parameters. Relying on extensive simulations, our proposed algorithm provably achieves better performance than IterML while maintaining comparable computational cost. Due to high dependency of this problem on the trade-off between the level of noise and the number of samples, we also investigate this compromise in order to achieve better reconstruction error metrics than IterML.

自2002年引入有限创新速率（FRI）信号以来，已提出了各种确定性和随机技术来估算FRI信号的创新参数，例如时间点 ${Ť}_{ķ}$ 和重量 $C_{ķ}$属于Diracs的有限数量的线性组合，来自其有噪样本，而与采样内核的类型无关。分析了专用于信号创新的贝叶斯方法，特别是Wein＆Srinivasan引入的IterML算法后，我们发现了一些局限性，尚有待改进。在本文中，我们提出了一种新颖的随机混合算法，该算法同时利用最大似然估计（MLE）和改进的粒子群优化（MPSO）来改善IterML的鲁棒性和估计参数的准确性。依靠广泛的仿真，我们提出的算法可证明比IterML具有更好的性能，同时保持可比的计算成本。