In this paper, we present a fast algorithm for filling unknown regions in an image using the strategy of exemplar-matching. Unlike the original exemplar-based method using exhaustive search, we decompose exemplars into the frequency coefficients and select fewer coefficients which are the most significant to evaluate the matching score. We have also developed a local gradient-based algorithm to fill the unknown pixels in a query image block. These two techniques bring the ability of input with varied dimensions to the fast query of similar image exemplars. The fast query is based upon a search-array data structure, and can be conducted very efficiently. Moreover, the evaluation of search-arrays runs in parallel maps well on the modern graphics hardware with graphics processing units (GPU). The functionality of the approach has been demonstrated by experimental results on real photographs.

We present a robust method to generate mesh surfaces from unoriented noisy points in this paper. The whole procedure consists of three steps. Firstly, the normal vectors at points are evaluated by a highly robust estimator which can fit surface corresponding to less than half of the data points and fit data with multi-structures. This benefits us with the ability to well reconstruct the normal vectors around sharp edges and corners. Meanwhile, clean point cloud equipped with piecewise normal is obtained by projecting points according to the robust fitting. Secondly, an error-minimized subsampling is applied to generate a well-sampled point cloud. Thirdly, a combinatorial approach is employed to reconstruct a triangular mesh connecting the down-sampled points, and a polygonal mesh which preserves sharp features is constructed by the dual-graph of triangular mesh. Parallelization method of the algorithm on a consumer PC using the architecture of GPU is also given.