Estimation of Depth Map Using Image Focus: A Scale-Space Approach for Shape Recovery

摘要

Laplacian-based derivatives used as a local focus measure to recover range information from an image stack have the undesirable effect of noise amplification, requiring good signal-to-noise ratios (SNRs) to work well. Such a requirement is challenged in practice by the relatively low SNRs achieved under classical phase contrast microscopy and the typically complex morphological structures of (unstained) live cells. This paper presents the results of our recent work on a new, multiscale approach to accurately estimate the focal depth of a monolayer cell culture populated with a moderately large number of live cells, whose boundaries were highly variable both in terms of size and shape. The algorithm was constructed in classical scale-space formalism which is characterised by an adaptive smoothing capability that offers optimal noise filtration/sensitivity and good localisation accuracy. Moreover, it provides a computationally scalable algorithm which not only obviates the need for additional heuristic procedures of global thresholding and (subsequent) interpolation of focus-measure values, but also generates as an integral part of the algorithm, a final range image/map that is demonstrably more realistic and, perceptually, more accurate.