Adaptive Inpainting Method Based On Image Retrieval With Global And Local Similarity

Inpainting means to recover the damaged parts of an image. For hand-paintedimages, the loss of part of the image reduction or specific targets to remove workfinished by skilled professional restoration, this way continues to the present. With theemergence of digital image, image restoration technology also become digital andintelligent, all kinds of image restoration methods develop very quickly. But the mainidea of most inpainting is to find a suitable patch from the original images in otherareas to repair the filled areas, leave the limitations that the found optimal patch in theoriginal image may not be able to match the surrounding background of the filledareas. In big data era, with the rapid development of Internet communication and datamining technology, real-time acquisition and processing of remote image data becomepossible, so the selection of image patch scope can be extended to the mass of theimage, which will undoubtedly lead to a big improvement of visual perception to therepair effects.In this article, we firstly use the scene and context similarity to query sampleimages which are matching with the original image, then we proposed an adaptiveblock matching repair method which proved more effective than other algorithms. Inorder to do so, we studied several exists work about image query and query-basedimage repair algorithms. The main works are as follows:First, we use a step-by-step image retrieval approach as an important foundation forrepairing. The first level is the global scene similarity based image retrieval, whichtake gist descriptor as the main characteristic of scene classification and retrieval thesemantic similar images from image database. The second level is the local contextsbased image retrieval, which takes textures around the filled region as the feature ofpartial context matching, and then continues to search for the most similar image (called sample image) in the first level results. Experiments shows, this two stageretrieval method can find superior pictures as the patch source than those globalsimilar query methods and local similar query methods, respectively.Second, proposed a regional adaptive block based transfer image restorationmethod. At the beginning, divide the repair area into several blocks, then looking forthe best matching patch in sample image to each block according to the principle ofgiving priority to large, and transferred the local color, tone and texture from originalimage to new added patches. Thus, we can avoid to the embarrassment that the repairarea is too large to find a suitable patch to fill. Compared with the whole piece offilling, this method can obtain better visual effect.