Research On Interaction And Composition Of Visual Media

With the rapid development of information technology and wide applications ofdigital media, visual media based on images and videos is increasingly becoming oneof the most important carriers of information acquisition, processing and dissemina-tion. In recent years, the interaction and composition of visual media is becoming a hotresearch topic of computer multimedia, its theory innovation and technological break-throughs will have a profound impact on the efficient use of visual media resources.It is expected to promote the development of digital content industry including film,animation, interactive entertainment and so on, which has the major significance ofresearch and applications.At present, in the interaction and composition of visual media, the bottlenecks ofinteractive editing, seamless composition, and harmonious environment still need to besolved. For the reasons, the paper focuses on the key techniques of interactive editing,image enhancement, and seamless composition. On the basis of efficient interactiveediting, it is emphasized to seek the harmonization of features such as boundary, sharp-ness, color and so on. In conclusion, in order to resolve the bottlenecks, we proposeseveral novel ideas on soft matting, edit propagation, saliency detection, sharpness en-hancement, motion deblurring and seamless cloning, which can improve the qualityand efficiency of the interaction and composition of visual media. Moreover, we alsointegrate these new techniques into a prototype system for editing and processing ofvisual media, which can further accelerate their promotion and applications. In thispaper, our main contribution includes:(1) In interactive editing, we first present hybrid matting using optimal con-straint parameter. The method employs the optimal constraint parameter to resolvethe smoothness assumption violation of closed-from matting and the inefficient confi- dence evaluation of robust matting, which can further improve the accuracy of objectextraction. Then, we present edit propagation for natural image using matting affinity.The method introduces the better matting affinity to reconstruct and optimize the en-ergy function, which can further improve the quality of edit propagation and achievethe efficient adjustment of appearances. Finally, we present video saliency detectionbased on dynamic contrast. The method defines an unified visual feature based oncolor, texture, and motion, and constructs its dynamic contrast and the region match-ing across frames, which can further improve the accuracy of video saliency detection.(2) In image enhancement, we first present edge-preserving sharpness enhance-ment in the gradient domain. Based on sharpness saliency representation, affinity-based gradient transformation and gradient-domain-based image reconstruction, themethod can resolve three kinds problems of image noise, unrealistic details and inco-herent enhancement in the traditional methods. Then, we present sharpness transfer forhigh-quality image composition. The method mainly includes four steps: sharpnessestimation, gradient transformation, image reconstruction and seamless composition.which can automatically and reasonably transfer the sharpness characteristics betweensource and target images. Finally, we present motion deblurring from a single imageusing sharpness enhancement. Based on sharpness enhancement, edge selection, ker-nel estimation and non-blind deconvolution, the method can accurately and efficientlyestimate the blurry kernel and yield the latent image.(3) In seamless composition, we first present seamless image composition withcoherent tone adjustment. On the basis of poisson cloning and image matting, themethod uses optimized parameters to effectively control boundary smudging and in-terpolation overflow, which can resolve color harmonization problem between twobackgrounds with distinct difference. Then, we present seamless video compositionbased on mean-value cloning, which mainly includes three key points: trimap prop-agation, smudging optimization and discoloring optimization. The method can notonly accurately relabel the user-drawn trimap between two neighboring frames butalso effectively eliminate the boundary smudging artifacts in mean-value cloning.