Visual Media Intelligent Processing Techniques By Structure Analyzing

In recent years,with the rapid development of electronic,network and multimedia techniques,digital cameras and smartphones become cheaper and more popular,and visual media-sharing websites like Facebook,YouTube and Flickr become more popular.The explosive visual media growth on the Internet has promoted the development of Internet business such as social network and e-commerce.Meanwhile,it causes great challenge in storage,computing,transmission and management of massive visual media.Therefore,the intelligent analysis and processing of visual media has become a hot research topic.Theoretical innovation and technological breakthroughs in this filed will have a profound impact on the efficient processing and effective use of visual media resources.It is expected to form business value for internet enterprises and exploit the potential applications of visual media.Research on visual media intelligent analysis and processing has great value in theory and practical significance.This thesis focuses on some hot topics of visual media intelligent analysis and processing,and studies some key issues about visual media recompression,intrinsic characteristics extraction,and face processing by analyzing visual media structure.Several novel ideas on massive internet images customized recompression,intrinsic image decomposition and face alignment are proposed and related algorithms are implemented.The main contributions of this paper are as follows:（1）A customized framework is presented to efficiently recompress massive internet images.The rapid growing number of images and their increasing resolution make a heavy burden on the device storage capacity and transmission bandwidth requirements.An objective image quality as-sessment,which is based on edges smoothness and texture distortion analysis,is used to predict the quality of recompressed images.Objective image quality assessment and subject image quali-ty assessment are combined in a unified framework which includes six components:initialization,image re-coding,quality measure,pipeline control,subjective evaluation,and custom service.This framework can provide customized massive image recompression for a variety of applications.It significantly saves storage and transmission bandwidth requirement and also improves user expe-rience.The proposed framework has been successfully applied to many commercial applications,such as web portals,e-commerce,online game,and so on.（2）A novel approach is proposed for automatic intrinsic images decomposition using near-L₀sparse optimization.It is an ill-posed problem to recover reflectance and shading components from a single input image.On the basis of the observation that the reflectance of natural objects is commonly piecewise constant,this paper formalizes this constraint on the entire reflectance im-age using the near-L₀sparse loss function that enforces the variation in reflectance images to be of high-frequency and sparse.An optimization formulation is proposed to consistently minimize the energy function transformed from the Maximum a Posteriori problem.This new sparsity con-straint significantly improves the quality of Retinex intrinsic images estimation.It also functions effectively by combining a class of global sparsity priors on reflectance.Experimental results on MIT benchmark dataset as well as various real-world images and synthetic images demonstrate the effectiveness and versatility of the proposed approach.（3）An intrinsic image decomposition method is proposed that uses multiscale measurements and sparsity.It generates and combines multiscale properties of chromaticity differences and in-tensity contrast.The key observation behind the proposed approach is that the estimation of image reflectance,which is neither a pixel-based property nor a region-based one,can be improved by using multiscale measurements of image content.It iteratively coarsens a graph reflecting the sim-ilarity between reflectance of neighboring pixels to form a pyramidal structure.In the process of pyramidal structure construction,we incorporate different-scale properties of the aggregates and modify the graph to reflect these different-scale measurements.Naturally,with the image coars-ening scheme,nodes in the higher level have larger support area that contain richer information to enable farther connections.What’s more,it uses a top-down manner to incorporate aggregated high-level information to facilitate intrinsic images decomposition.This problem is formulated through energy minimization which can be solved efficiently within a few iterations.Experimental results have been conducted on the MIT benchmark dataset,Intrinsic Image in the Wild dataset and various natural images,demonstrating the effectiveness of the proposed approach.（4）It proposed that a face alignment method based on L₁sparsity constraint.Facial images,which are very active on the internet,are the majority of massive internet images.Facial landmark-s contain geometrical structure of face and abundant semantic information.Effective and robust face alignment really benefits applications such as large-scale face image retrieval and visual me-dia management and exploration.Based on the explicit shape regression framework,this paper proposes a face ratio regression method which can help to solve the problem of training face align-ment model with different scales samples.Moreover,an L₁punishment is added on the regression variables which enforces the regression shape to be sparsity.This sparsity constraint makes the regression model more stable and helps to reduce the size of the trained model.Since the proposed method jointly regresses the entire face shape,the geometric structure of landmarks is implicit en-coded in the model.Experimental results on some related datasets demonstrate the accuracy and efficiency of the proposed face alignment method.