An unidirectional Criminisi Algorithm for DIBR-synthesized images

Virtual free viewpoint images in a 3D space can be synthesized by Depth-Image Based Rendering (DIBR). But DIBR images contain holes which are occluded background regions —that requires holes should be filled with background textures. To address this problem, an unidirectional Criminisi algorithm is proposed in this paper. First, the hole-contour is extracted by using Laplacian operator. Based on view-switch direction, a novel filling line is extracted from hole-contour. Second, candidate patches are selected with the central pixel on the filling line. This ensures the patches with known background textures can be filled first. Then the best to-be-filled patch is selected from candidates by using Criminisi method and depth limitation. Third, in template matching process, SAD method and depth limitation are used for seeking the best source patch. Experimental results show the proposed algorithm can outperform the hole-filling by extending the nearest background textures into holes. It improves the image quality by 1.35dB and 1.17dB in PSNR compared to Criminisi Algorithm and Criminisi-Based-on-Depth Algorithm respectively.