Research On Semantic Image Retargeting

Image retargeting is an important research area of computer graphics and imageprocessing. Previous methods scale images into target size, but fail to keep the imagein a semantic way. Images maybe shrink badly and main parts of images will bedistorted during the resizing process. At the other hands, some approaches only retainthe important parts and discards the surrounding to fit the display resolution, thisusually discard large number of background information. The retargeting methodswithout semantic feature usually lead to user’s misunderstanding. Image retargetingbased on semantic recognizing considers image content in the process of imageresizing, which is now one of the research focuses.This thesis first surveys the existing works on image retargeting. The surveyincludes theoretical basis, technical characteristics, related work, difficulties,development and applications. Based on three kinds of specific images (texture,nature image and image with symmetry), the thesis proposes different retargetingmethods.For natural images, seam carving is an effective operator supportingcontent-aware resizing for both image reduction and expansion. However, repeatedseam removing and inserting processes lead to excessively distortion image whenimposed on seam insertion and removal operations or the other way around. Byconsidering the relationship between seams removing and inserting processes, wepresent an ameliorated energy function to minimize aliasing.“Forward Energy” isonly an effective improvement to image reduction. Moreover, the thesis propose anovel”Visual Points” structure which distinguishes the “Forward Energy” of seaminsertion from that of seam removal, and improves seam insertion operations greatly.The thesis introduces texture synthesis method to texture image retargeting.Time consumption and quality are two main concerns to texture synthesis algorithm.A wavelet based texture optimization approach is proposed in this paper. Twomulti-resolution texture pyramids are used: an input pyramid built by the wavelettransform of the exemplars and an output pyramid reconstruction from the inversewavelet transform. In the step of nearest neighborhood searching, wavelet coefficientis integrated to estimate neighborhoods’ distance, instead of RGB and other channels.Because the wavelet transform is reversible and nondestructive, this strategy does notdebase quality. Images with symmetry exist everywhere. Symmetry summarization can resize atranslational symmetric image while maintaining the semantics quite well. The thesispresents a novel retargeting method that allows arbitrarily resizing the ratio ofrotational symmetric images as well as preserving their symmetric structure andseamlessness. First, we detect the rotational symmetric area, and convert the area intolinearly translation symmetric form. Then, we construct a texture feature map andextract potential symmetric cells through symmetry group analysis. The resizingprocess is finally achieved by increasing or removing proper-measured symmetriccells. Both the experimental results and user study demonstrate superior performanceof our algorithm compared with other existing methods.At last, some examples are illustrated applying the proposed algorithm ondifferent kinds of images, which concludes the thesis with discussions of the digitalresults and limitations of the presented method.