基于粒子滤波的人脸图像超分辨率重建方法

将人脸图像超分辨率重建描述为人脸混合模型的纹理和位置参数的贝叶斯概率估计问题,将超分辨率重建的图像配准和像素融合这两个过程置于统一的概率估计框架下,并利用基于粒子滤波的参数估计算法,同时估计纹理和位置参数,从而实现人脸图像的超分辨率重建.包含灰度和位置两种先验信息的人脸混合模型,同时用于超分辨率重建的两个过程中,提高了图像配准精度和重建算法的性能,避免了通常方法在获得准确鲁棒的运动场估计时需要清晰的高分辨图像,而获得清晰的高分辨图像时又需要准确鲁棒运动场估计的困境.正面人脸合成序列图像实验结果表明,该方法获得的重建结果较为理想.

Super-Resolution Reconstruction for Face Images Based on Particle Filters Method

Super-Resolution (SR) reconstruction is posed as a Bayesian estimation of the location and appearance parameters of a face model. Image registration and image fusion, the two steps for SR reconstruction, are combined into one unified probabilistic framework, in which the prior information about facial appearance and gray from the face model is incorporated into both of the steps. In addition, a particle filter based algorithm is proposed to achieve the estimation, i.e. SR reconstruction. The proposed approach avoids the inherent dilemma of the most traditional methods, in which it demands a high-resolution image to get an accurate and robust estimation of the motion field, while reconstructing a high-resolution image requires the accurate and robust estimation of motion field. Experiments performed on synthesized frontal face sequences show that the proposed approach gains superior performance both in registration and reconstruction.