Shape from focus using kernel regression

In conventional focus measures, focus values are locally aggregated to suppress the noise and to obtain better depth maps. However, this enlarges the difference between focus values of two consecutive frames which results in inaccurate shape. In this paper, we propose a nonparametric approach for 3D shape from image focus by applying an unsupervised formulation of kernel regression estimate. The focus volume is obtained through a focus measure and then Nadaraya and Watson Estimate (NWE) is applied to each frame. The depth is then computed by finding the frame number which maximizes the focus value. The kernel regression is again applied on depth values to obtain an accurate 3D shape. The proposed approach is experimented using synthetic and real image sequences. The results demonstrate the effectiveness of the proposed approach.