In shape-from-focus (SFF) methods, a single focus measure is used to compute the focus volume. However, it seems that a single focus measure operator is not capable of computing accurate focus values for the images of diverse types of object shapes. Furthermore, most of the SFF methods try to improve the depth map without considering any additional structural or prior information. Consequently, the extracted shape of the object might lack important details. In this work, we address these problems and suggest a method in which depth hypotheses are combined for a more accurate 3D shape through 3D weighted least squares. First, depth hypotheses are obtained by applying a number of focus operators. Then, structural prior or guidance volume is extracted from the focus measure volumes. Finally, a 3D weighted least squares optimization technique is applied to the depth hypothesis volume, where weights are computed from the guidance volume. Thus, by inducing structural prior, an improved resultant depth map is obtained. The proposed method was tested using various image sequences of synthetic and microscopic real objects. Experimental results and comparative analysis demonstrated the effectiveness of the proposed method.

Shape from focus (SFF) is a passive optical technique that reconstructs object shape from a sequence of image taken at different focus levels. In SFF techniques, computing focus measurement for each pixel in the image sequence, through a focus measure operator, is the fundamental step. Commonly used focus measure operators compute focus quality in Cartesian space and suffer from erroneous focus quality and lack in robustness. Thus, they provide erroneous depth maps. In this paper, we introduce a new focus measure operator that computes focus quality in log-polar transform (LPT) Properties of LPT, such as biological inspiration, data selection, and edge invariance, enable computation of better focus quality in the presence of noise. Moreover, instead of using a fixed patch of the image, we suggest the use of an adaptive window. The focus quality is assessed by computing variation in LPT. The effectiveness of the proposed technique is evaluated by conducting experiments using image sequences of different simulated and real objects. The comparative analysis shows that the proposed method is robust and effective in the presence of various types of noise.