The Research And Applications On Several Issues Of Internet Visual Media

. With the rapid development of the Internet, more and more multimedia resources such as images, videos and texts are uploaded on the Internet by users. Among these modalities, images and videos are the most active information carriers as they are capable of providing visual information intuitively and efficiently. Apart from the large-scale visual information, additional information from other extensive modalities, e.g. tagging, comments and user preferences, can be also leveraged to facilitate applications with respect to multimedia content analysis and understanding. This research usually involves multiple fields such as computer graphics, computer vision and machine learning, and its aim is to better develop novel methods and systems for user-oriented intelligent applications.In this context, we aim to develop several novel systems for intelligently representing and analyzing multimedia contents. The studies in this dissertation consist of three parts: multi-modality artistic image rendering, face stylization rendering and video animation for shadow puppetry, efficient visual features extraction and classification for furniture style. The main contents and contributions of this dissertation are briefly stated as follows:1. We propose a method for multi-modality image rendering by fusing information from text and implement Picwords system.Image-based artistic rendering is an image processing technology that focuses on image abstraction and stylization. We propose a novel artistic image rendering method to extend the semantic information of original image by integrating two modalities, i.e. image and text, into one target picture seamlessly. The source image information corresponds to the low frequency signal and represented as holistic information. The text information corresponds to the high frequency signal and is represented as detailed information. Based on this method, we implement a multi-modal image rendering system Picwords. This system can fuse the picture and the keywords into one target Piwords image and rerank the keywords automatically. The output of the system reserves the entirety visual effect and enriches the semantic information of the input image. Pic Words also has great market potentials of poster design, advertisement and social networks.2. We propose face artistic rendering and interactive animation techniques for shadow puppetry, and present a novel digitized eHeritage system of puppetry.To preserve the precious traditional heritage Chinese shadow puppetry, we propose an eHeritage system for shadow puppetry, including a creator module and a manipulator module, which allows personalized character creation and automated manipulation based on the images and videos of shadow puppetry on the Internet. For the creator module, we utilize the face artistic rendering method to generate the specialized puppet head according to the input natural facial image. For the manipulator module, the interactive animation is implemented to generate puppetry motion sequence automatically by providing manipulating scripts. The puppetry eHeritage system can simplify the operation of puppetry and keep the characteristics of traditional Chinese shadow puppetry, which demonstrates the effectiveness of preservation and dissemination of this system for shadow puppetry.3. We propose a method for furniture style visual classification based on the handcrafted features and deep learning features.Since furniture style describes the discriminative appearance characteristics of furniture and plays an important role in real-world indoor decoration, furniture style analysis has various potential applications for commercial advertisement and recommendation. Different from traditional object classification, furniture style classification targets at classifying different appearance of furniture in details such as materials, color, texture, size and lines. To pursue efficient representation of furniture style, we firstly construct a novel dataset of furniture styles, which is also the first image dataset for furniture style research. Then we compare the classification performance of traditional handcrafted classification and deep learning classification. We also propose a multi-scale convolutional feature. Finally, we combine the pipeline of handcrafted classification based on the deep learning classification at different levels and achieve a classification accuracy of70%for a16-class classification problem.