Zhuang Yue-ting

An iris recognition method based on multi-orientation features and Non-symmetrical SVM *

A new iris feature extraction approach using both spatial and frequency domain is presented. Steerable pyramid is adopted to get the orientation information on iris images. The feature sequence is extracted on each sub-image and used to train Support Vector Machine (SVM) as iris classifiers. SVM has drawn great interest recently as one of the best classifiers in machine learning, although there is a problem in the use of traditional SVM for iris recognition. It cannot treat False Accept and False Reject differently with different security requirements. Therefore, a new kind of SVM called Non-symmetrical SVM is presented to classify the iris features. Experimental data shows that Non-symmetrical SVM can satisfy various security requirements in iris recognition applications. Feature sequence combined with spatial and frequency domain represents the variation details of the iris patterns properly. The results in this study demonstrate the potential of our new approach, and show that it performs more satis- factorily when compared to former algorithms.

Digital libraries: A testbed for multimedia technology

A distinguishing feature of a digital library is that it has Terabyte volumes of multimedia resources. One challenge for researchers in the field of multimedia is to find a testbed for showing the potentials of multimedia technologies such as video summarization, semantic annotation, multimedia cross indexing and retrieval, and etc. Deeper research and wider applications of digital libraries revealed their indispensable role as testbed for multimedia technologies. This paper presents challenging issues of some key techniques used in digital libraries and their specific needs for multimedia technologies.

Ensemble learning HMM for motion recognition and retrieval by Isomap dimension reduction

Along with the development of motion capture technique, more and more 3D motion databases become available. In this paper, a novel approach is presented for motion recognition and retrieval based on ensemble HMM (hidden Markov model) learning. Due to the high dimensionality of motion’s features, Isomap nonlinear dimension reduction is used for training data of ensemble HMM learning. For handling new motion data, Isomap is generalized based on the estimation of underlying eigenfunctions. Then each action class is learned with one HMM. Since ensemble learning can effectively enhance supervised learning, ensembles of weak HMM learners are built. Experiment results showed that the approaches are effective for motion data recognition and retrieval.

Image retrieval system for Web: Webscope-CBIR

With the widespread usage of images on Web, the retrieval need for images has been increasing rapidly. But traditional text retrieval system is not fit for the application of image retrieval. We therefore designed Webscope-CBIR, an image retrieval system for the Web, which utilizes content based retrieval to find similar images by comparing visual features. It also uses HTML file and the procedure of relevance feedback to dynamically extract image semantics. The Webscope-CBIR system integrates visual and semantic features of images to improve the preciseness of retrieval.

Translating EXPRESS language model into C language model

EXPRESS is a powerful object-oriented data model descriptive language and independent of any platform. However, it is a kind of descriptive language rather than a programming language. This brings difficulty to implement the EXPRESS data model on a computer and it also undoubtedly brings a big problem with the STEP data model processing so that we have to translate EXPRESS language model into a certain programming language model. In this paper, an approach is presented to translate the EXPRESS language model into the C language model. In addition, this paper proves the validity of this translation theoretically.EXPRESS is a powerful object-oriented data model descriptive language and independent of any platform. However, it is a kind of descriptive language rather than a programming language. This brings difficulty to implement the EXPRESS data model on a computer and it also undoubtedly brings a big problem with the STEP data model processing so that we have to translate EXPRESS language model into a certain programming language model. In this paper, an approach is presented to translate the EXPRESS language model into the C language model. In addition, this paper proves the validity of this translation theoretically.

Using CONDENSATION Tracking to Recover Stroke Order of Chinese Calligraphic Handwritings with CCM

Stroke order is an intrinsic writing feature which is missing in off-line handwriting. Recovering stroke order from off-line handwriting is an important task by which the visualization of writing process is possible. Chinese Calligraphic handwriting is some different with western handwriting. Traditional method with assumption of the minimization of curvature between two adjacent line segments is not applicable for Chinese Calligraphic handwriting. We propose a new framework named Chinese Calligraphic Model (CCM). With this framework CONDENSATION algorithm - conditional density propagation over time is performed to find the best matching model in model database with the input pattern to recover stroke order. Experiments results show our method is a promising way to recover stroke order from Chinese Calligraphic handwriting based on CCM by CONDENSATION tracking.

Video segmentation using Maximum Entropy Model

Detecting objects of interest from a video sequence is a fundamental and critical task in automated visual surveillance. Most current approaches only focus on discriminating moving objects by background subtraction whether or not the objects of interest can be moving or stationary. In this paper, we propose layers segmentation to detect both moving and stationary target objects from surveillance video. We extend the Maximum Entropy (ME) statistical model to segment layers with features, which are collected by constructing a codebook with a set of codewords for each pixel. We also indicate how the training models are used for the discrimination of target objects in surveillance video. Our experimental results are presented in terms of the success rate and the segmenting precision.

A Novel Incremental Mining Algorithm of Frequent Patterns for Web Usage Mining

Because data warehouse is frequently changing, incremental data leads to old knowledge which is mined formerly unavailable. In order to maintain the discovered knowledge and patterns dynamically, this study presents a novel algorithm updating for global frequent patterns-IPARUC. A rapid clustering method is introduced to divide database into n parts in IPARUC firstly, where the data are similar in the same part. Then, the nodes in the tree are adjusted dynamically in inserting process by “pruning and laying back” to keep the frequency descending order so that they can be shared to approaching optimization. Finally local frequent itemsets mined from each local dataset are merged into global frequent itemsets. The results of experimental study are very encouraging. It is obvious from experiment that IPARUC is more effective and efficient than other two contrastive methods. Furthermore, there is significant application potential to a prototype of Web log Analyzer in web usage mining that can help us to discover useful knowledge effectively, even help managers making decision.

Video motion capture in VBA—Video-based animation

Computer vision has very wide application in human motion capture research. This paper proposes a new approach to do motion capture in video. It is composed of image sequence based tracking of human feature points and the reconstruction of the three-dimension(3D) motion skeleton. First, every part of the human body from top to bottom is tracked on the basis of a human model. The image difference and a morph-block similarity algorithm based on subpixels are used. Then camera calibration is done using the line correspondences between the 3D model and the image. Finally the 3D motion skeleton is established by use of the model knowledge. This approach does not aim at a given mode of human motion. Rather, it analyzes large scale motion from frame to frame in complex, variational background, and sets up a 3D motion skeleton in the perspective projection. The experiment results are presented at the end of the paper.

A retrospective event detection method in news video

In this work we present a probabilistic learning approach to model video news story for retrospective event detection (RED). In this approach, both content and time information on a news video is utilized to transcribe the news story into terms, which are divided into classes by their semantics. Then a probabilistic model, composed of sub-models corresponding to the semantic classes respectively, is proposed. The model’s parameters are estimated by EM algorithm. Experiments showed that the proposed approach has better detection resolution.