Catálogo de publicaciones - libros

High-Definition (HD) videos often contain rich details as well as large homogeneous regions. To exploit such a property, Variable Block-size Transforms (VBT) should be in place so that transform block size can adapt to local activities. In this paper, we propose a 16× 16 Integer Cosine Transform (ICT) for HD video coding, which is simple and efficient. This 16×16 ICT is integrated into the AVS Zengqiang Profile and used adaptively as an alternative to the 8×8 ICT. Experimental results show that 16×16 transform can be a very efficient coding tool especially for HD video coding.

In this paper a novel distributed video coding scheme was proposed based on Heegard-Berger coding theorem, rather than Wyner-Ziv theorem. The main advantage of HB coding is that the decoder can still decode and output a coarse reconstruction, even if side information degrade or absent. And if side information present or upgrade at decoder, a better reconstruction can be achieved. This robust feature can solve the problem lies in Wyner-Ziv video coding that the encoder can hardly decide the bit rate because rate-distortion was affected by the side information known only at the decoder. This feature also leaded to our HB video coding scheme with 2 decoding level of which we first reconstruct a coarse reconstruction frame without side information, and do motion search in previous reconstructed frame to find side information, then reconstruct a fine reconstruction frame through HB decoding again, with side information available.

MPEG-21 BSDL offers a solution for exposing the structure of a binary media resource as an XML description, and for the generation of a tailored media resource using a transformed XML description. The main contribution of this paper is the introduction of a real-time work flow for the XML-driven adaptation of H.264/AVC bitstreams in the temporal domain. This real-time approach, which is in line with the vision of MPEG-21 BSDL, is made possible by two key technologies: BFlavor (BSDL + XFlavor) for the efficient generation of XML descriptions and Streaming Transformations for XML (STX) for the efficient transformation of these descriptions. Our work flow is validated in several applications, all using H.264/AVC bitstreams: the exploitation and emulation of temporal scalability, as well as the creation of video skims using key frame selection. Special attention is paid to the deployment of hierarchical B pictures and to the use of placeholder slices for synchronization purposes. Extensive performance data are also provided.

Not much has been done in utilizing the available information at the decoder to optimize the decoding performance of watermarking systems. This paper focuses on analyzing different decoding methods, namely, Minimum Distance, Maximum Likelihood and Maximum decoding given varying information at the decoder in the blind detection context. Specifically, we propose to employ Markov random fields to model the prior information given the embedded message is a structured logo. The application of these decoding methods in Quantization Index Modulation systems shows that the decoding performance can be improved by Maximum Likelihood decoding that exploits the property of the attack and Maximum decoding that utilizes the modeled prior information in addition to the property of the attack.

Traditional 2D texture synthesis methods mainly focus on seamlessly generating a big size texture, with coherent texture direction and homogeneous texture scale, from an input sample. This paper presents a method of synthesizing texture with variational direction and scale on arbitrary planar region. The user first decomposes the interest region into a set of triangles, on which a vector field is subsequently specified for controlling the direction and scale of the synthesized texture. The variational texture direction and scale are achieved by mapping a suitable texture patch found in the sample via matching a check-mask, which is rotated and zoomed according to the vector field in advance. To account for the texture discontinuity induced by not well matching or different texture directions/scales between adjacent triangles, a feature based boundary optimization technique is further developed. Experimental results show the satisfactory synthesis results.

The four most important issues in content-based image retrieval (CBIR) are how to extract features from an image, how to represent these features, how to search the images similar to the query image based on these features as fast as we can and how to perform relevance feedback. This paper mainly concerns the third problem. The traditional features such as color, shape and texture are extracted offline from all images in the database to compose a feature database, each element being a feature vector. The “linear scaling to unit variance” normalization method is used to equalize each dimension of the feature vector. A fast search method named equal-average K nearest neighbor search (EKNNS) is then used to find the first K nearest neighbors of the query feature vector as soon as possible based on the squared Euclidean distortion measure. Experimental results show that the proposed retrieval method can largely speed up the retrieval process, especially for large database and high feature vector dimension.

The universal recognition that it is critical to improve the performance of existing systems and protocols with the understanding to practical service experiences motivates us to discuss this issue in the context of peer-to-peer (P2P) streaming. With the benefit of both practical traces from traditional client-server (C/S) service systems and logs from P2P live broadcasting system, in this paper we first characterize end user behaviors in terms of online duration and reveal the statistically positive correlation between elapsed online duration and expected remaining online time. Then we explore the feasibility to improve the quality of streaming service over P2P networks by proposing Low Disruption Tree Construction (LDTC) algorithm to take the online duration information into account when peers self-organize into the service overlay. The experiment results show that LDTC could achieve higher stability of video date delivery tree and in turn improve the quality of streaming service.

Activity recognition is one of the most challenging problems in the high-level computer vision field. In this paper, we present a novel approach to interacting activity recognition based on dynamic Bayesian network (DBN). In this approach the features representing the human activities are divided into two classes: global features and local features, which are on two different spatial scales. To model and recognize human interacting activities, we propose a hierarchical durational-state DBN model (HDS-DBN). HDS-DBN combines the global features with local ones organically and reveals structure of interacting activities well. The effectiveness of this approach is demonstrated by experiments.

In the area of image retrieval, search engines are tender to retrieve images that are most relevant to the users’ queries. Nevertheless, in most cases, queries cannot be represented just by several query words. Therefore, it is necessary to provide relevant retrieval results with broad topic-coverage to meet the users’ ambiguous needs. In this paper, a re-ranking method based on topic coverage analysis is proposed to perform the refinement of retrieval results. A graph called Topic Coverage Graph (TCG) is constructed to model the degree of mutual topic coverage among images. Then, Topic Richness Score (TRS), which is calculated based on TCG, is used to measure the importance of each image in improving the topic coverage of image retrieval results. Experimental results on over 20,000 images demonstrate that our proposed approach is effective in improving the topic coverage of retrieval results without loss of relevance.

This paper presents a hierarchical motion-compensated frame interpolation (HMCFI) algorithm based on the pyramid structure for high-quality video reconstruction. Conversion between images having different frame rates produces motion jitter and blurring near moving object boundaries. To reduce degradation in video quality, the proposed algorithm performs motion estimation (ME) and motion-compensated frame interpolation (MCFI) at each level of the Gaussian/Laplacian image pyramids. In experiments, the frame rate of the progressive video sequence is up-converted by a factor of two and the performance of the proposed HMCFI algorithm is compared with that of conventional frame interpolation methods.