Research On Sample Based Image Completion

Image is one of the most significant information carriers. Imagecompletion, aiming at improvement of visual effect, has aroused wideconcern and attracted intensive study. In the process of image acquisition,compression, transmitting, storage and image editing, the missing ordegradation of part of image information will heavily damage image’svisual effect. Image completion technique employs the known imageinformation to compute, predict and recover the lost data. After years ofstudy and research, image completion algorithms have achieved prettygood restoration results for small range image data lost problem, texturesynthesis and object removal problems. However, for large region imagecompletion question, state of art image completion techniques cannotrecover degraded image to be relatively visually acceptable. This paperanalyzes large region image information lost problem and summarizesstate of art image completion algorithms. Based on the analysis, novelsample based image completion algorithms are proposed to improve imagecompletion effect. The achievements are demonstrated as below:First of all, a sample based image structure reconstruction algorithmis presented. Structure is very significant in recovering the degraded image.In order to reconstruct the structure of degraded image, the proposedalgorithm extracts the foreground structure from sample image and makesuse of boundary matching and Bezier to link sample structure and structureof source image. Based on proposed structure reconstruction algorithm,image completion effect can be improved because of the reconstructedstructure helps to hold image’s real content.Secondly, after reconstruction of image structure, this paper proposesa novel sample-based image synthesis algorithm. In large range image data lost problem, algorithms using source image information to complete theunknown cannot keep the image consistency because there is so littleuseful pixels in the left region of source image. In this paper, textures ofboth sample image and source image are employed to restore target region.At beginning, algorithm computes the correlation between the missingregion and the sample. Then, it searches in the sample image and sourceimage to find the best image patches. According to the sample correlationparameter, combination of sample patch and source patch is put in thetarget patch at last. Experiment results show that proposed algorithmsuccessfully recovers the content of damaged image, while maintaining theimage consistency.Thirdly, this paper modifies the web-based image completionalgorithm presented by Hays, and presents a novel image completiontechnique based on sample index, along with the sample-based imagecompletion algorithm. Hays presented his algorithm which index sampleimage in web-based image database, but he just used SSD as sample localmatching metric and graph cut technique to restore image. In this paper,gradient orientation is employed to establish local matching metric in orderto choose the most structurally similar sample. After using sample-basedimage completion algorithm, the experimental results are better.