一种新的正则化粒子滤波图像恢复算法

Image Restoration Based on a New Regularized Particle Filter

引入粒子滤波对解决非线性非高斯模型的优良特性,将一种新的正则化粒子滤波算法(regularized particle filter)应用到混合噪声和乘性噪声图像恢复中.由于采样重要性重采样(SIR)方法在重采样时没有考虑观测量而引入误差,本文方法从后验连续分布中采样,引入观测量进而减少误差,同时将易实现的累积分布函数和正则化重采样步骤融合,进一步使粒子方差最小化,解决粒子衰竭问题,缓解退化现象.通过对具有混合噪声图像以及医学乘性噪声图像的恢复效果表明了该算法的有效性,且与小波阈值法和SIR粒子滤波法对比显示了其优越性.

It is known that particle filter has the superior characteristics of solving non-linear non-Gaussian problems.A new regularized particle filter is proposed to improve the quality of image restoration with mixed or multiplicative noise in this paper.To reduce the error of sample importance re-sampling（SIR） particle filter,which comes from the neglect of measuring data when re-sampling,the posterior continuous distribution sample is adopted.The combination of the cumulative distribution function（CDF） and regularized re-sampling step makes the algorithm have the advantages of minimized variance,alleviating degradation and escaping from the exhaustion of particle.The proposed method has been applied to the mixed noisy and medical multiplicative noisy image restoration.The results show the effectiveness of the algorithm,and demonstrate its superiority,compared with wavelet threshold shrink method and SIR particle filter method.