Patched Image Inpainting Based on Texture Segmentation and Structure Reconstruction

• Main Authors: Tung-lin Wang, 王苳霖
• Other Authors: Nai-jian Wang
• Format: Others
• Language: zh-TW
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/70045826318076857286

碩士 === 國立臺灣科技大學 === 電機工程系 === 97 === In past years, several researchers proposed various image inpainting algorithms for removing objects. There are two popular inpainting algorithms among previous works. The first one is the PDE-based image inpainting, which propagates image in-formation into target regions via diffusion. The main drawback of PDE-based image inpainting is the result image with some blurring effect in target regions when the area is large. The second one is the exemplar-based image inpainting, which copies image information to target regions by texture comparison, and can restore large occluding objects, but exemplar-based image inpainting cannot correctly reconstruct non-linear image structures, and furthermore, it causes texture sampling error frequently. The main contribution of this thesis is to improve the drawbacks of exemplar-based image inpainting. We introduce the same concept of exemplar-based image inpainting into our algorithm, then preprocess target image with texture segmentation and structure reconstruction. By our proposed algorithm, the time consumption of searching texture is reduced and the accuracy of texture sampling by texture segmentation increases. In structure reconstruction, we analyze all structures in the image except the target re-gions, and reconstruct image structures with non-linear curves in the target regions. Our experiments show that the reconstructed structures and textures of result image are favorably compared to those of conventional exemplar-based image inpainting techniques.