Research Of Image Editing Algorithms Based On Internet Vision

With the development of computer vision,image seems more as an important information medium.People get to know and understand the world by referring to the Internet which is known of vast richness information in images,thus the understanding of one image various from people to people.And image editing in turn satisfies the need of that.Imgae editing,combining needs of people and content analysis of image itself,generates one reasonable result by reasonable editing information.In most cases,reliable editing information means a good editing result.Thus what can we do with the editing information,how to find them and how can we have a visual appealing results by them,all of them matter the process of image editing.The editing information from different sources holds different advangates.The editing information from the image itself has relative high accuracy while less options.The alterative options are improved in a collection of image database but falls in high accuracy.Taking tha advantages of Big-Data,the Internet can offer both accuracy and option choices of editing information.In this way how to retrieve the needed editing information and make the full use of it in image editing worthy of further study.The main research contents of this thesis include*:(1)A scene images inpainting algorithm based on Gist is proposed in this thesis due to the lassness of inpainting,which is one of the editing process,whose information existed in one image itself as well the undesired target region to be inpainted in most traditional inpainting algorithms.Firstly,‘One-click' process by user clicking once to obtain the desired target region based on a pre-segmentation based on SLIC(Simple Linear Iterative Clustering),which the target region hold both satisfaction of user and accuracy of the target region.Secondly,a Gist feature is used together with a multi-feature constrain helps retrieve the candidates by similarity.After that a global sliding window is used in the best candidate to scan for the most approate patch with the target region.In the end a Poisson blending is used to smooth the boundary between the patch and the target region,this helps remove blurring and sharpening along with the copy of patch into the target region.Experiments show that the proposed algorithm achieves good visual effect in many kinds of scene image inpainting.(2)By the study of the inpainting process,how inpainting information resembles the target images effects the result.Therefore,an Internet vision image editing algorithm based on spatiotemporal feature is proposed in this thesis.This contains two major procures.One is a standout Internet vision image retrieval and another is an image editing process of both replacement as well as inpaintng based on the retrieval results.The retrieval process contains an interactive process of local offline database and Internet online search helping improve the accuracy of image retrieval result.By finding the inner relationship between the spatio-temporal feature and the circumstance of one particular image,a spatio-temporal feature is used as a new attribute in image retrieval.Firstly,a spatio-temporal feature together with Bo W(Bag-of-Words)feature is used as the classification feature of the local offline database images.And thus corresponding labels are assigned to each category.Secondly,calculating the corresponding features in one retrieval image and using TF-IDF(Term Fervency-Inverse Document Frequency)as a category label assignment as well as spatio-temporal label by the calculating the differences.With the labels obtained,an Internet online retrieval process is carried out.The unfiltered results is evaluated in four aspecsts by the using of Internet visual feature to have accuracy results.Finally,a pre-segmentation based on Robust matting as well as a Bayesian Probability is used to obtain a well-segmented forground.In this way an Alpha blending is used to blend the forground with the retrieval result as a new background.In order to have a consistence of the new background and the former foreground,a transmission of both light and color is performed.Experiments show that the proposed algorithm has good accuracy in image retrieval thus guarantee the effectiveness of the editing information,which in turn improves the visual appealing of image editing results.