Research Of Image Completion Based On Structure Information And Patch Statistics

Image completion, or image inpainting as an emerging field of image processing, is a task to fill the missing region in an image in a visually pleasant way. Image completion has a wide range of applications such as image editing, heritage restoration,image coding and transmission, and thus has great significance and research value. As existing novel inpainting methods cannot recover the salient structure plausibly when the image missing part is large or the important structure is lost, in this thesis, structure reconstruction and new priority assignment scheme are proposed to improve the image inpainting effect. The main contributions of this thesis are as follows:(1) An image completion algorithm based on structure reconstruction and constraint is presented. For important structure in large area degraded image cannot be repaired commendably by current inpainting algorithms, structure reconstruction is employed to preserve and extend the original structure into the missing region of the image. To extract the main interrupted structure for reconstruction, image segmentation is applied.And the main broken structure is reconstructed by Euler spiral which satisfies energy minimization. The reconstructed structure is used as the constraint to guide the texture propagation until the completion finish. The proposed algorithm preserves structure continuity while avoids texture inconsistency effectively, and obtains satisfactory visual result.(2) An image completion algorithm based on patch gradient statistics is presented.As the isophote-driven priority often leads to structure discontinuity and block effect,the internal feature of the image and the statistical characteristics between image patches are analyzed to get a more reasonable priority assignment scheme. Gradient reflects the image frequency variation and the structure extension direction, so the gradient volatility between the neighboring source patches is used to determine the filling-in order. The proposed algorithm ensures a better maintenance of structure coherence, and improves the image inpainting quality.(3) A matching method that combines color and gradient information is proposed to measure the similarity between two patches. Using only color information to measure similarity of patches often leads to false match that the color are similar but the structure and texture are different. Due gradient information reflects the pixels variations and direction information, the color information and the gradient information are combinedto measure the similarity of two patches. By using the proposed matching method that considers both color and gradient information, the matching patch that searched for target patch is more similar in visual and the directions are basically the same. Moreover,the inpainting result is more plausible.