Deinterlacing with Motion-Compensated Anisotropic Diffusion

Abstract

Abstract. We present a novel deinterlacing scheme that makes consequent use of discontinuity-preserving partial differential equations (PDEs). It combines the accuracy of recent variational motion estimation techniques with the directional interpolation qualities of anisotropic diffusion filters. Our algorithm proceeds in three steps: First, we interpolate the interlaced images by means of a spatial edge enhancing diffusion process (EED). Then we apply the variational optic flow technique of Brox et al. (2004) in order to obtain a precise interframe registration. Finally we use a spatiotemporal generalisation of EED for motion-compensated inpainting of the missing data in the original sequence. Experiments demonstrate that the proposed method outperforms not only classical deinterlacing schemes, but also a recent PDE-based approach.

Motion Compensated Frame Interpolation with a Symmetric Optical Flow Constraint

“…[14] and [11]) or by edge enhancing diffusion [18]. The latter has been shown to have very good interpolation properties, and has been successfully used in image compression [19] and for motion compensated deinterlacing [20]. To improve reconstruction quality in terms of e.g.…”

A Variational Method for Dejittering Large Fluorescence Line Scanner Images

“…Inspired from optical flow models, the estimation of the underlying motion have been recently used in state-of-the-art deinterlacing methods. Most of these motion-estimation-based methods are defined in a variational framework with TV regularization [14,15,16].…”

“…Irrespective of deinterlacing or dejittering algorithms, we can classify the majority of them into three categories : (a) simultaneous approach as [4,16] which consists in estimating jointly the displacement and the jitter-free original image ; (b) displacement estimation methods which represent the majority part on the literature as in [2,1,3,5,7,8,9,14] ; (c) image correction without estimating the displacement as in [15,10,13] which is mainly based on the BV image model and variational methods. However, there are problems which are not fully addressed such as large and non-integer displacement, over-fitting, or over-smoothing.…”

“…This problem can be solved point-wise, since there are no spatial dependency term for w in (14). Therefore, the optimal values for w i,j at every point (i, j) can be simply computed by an exhaustive search in the discrete set of possible displacements…”