In this talk, I will mainly focus on the introduction of classic graph-based interactive image segmentation approaches. Different from automatic segmentation, interactive segmentation involves user interaction as the prior knowledge to effectively control the process of segmentation. Graph theory is commonly used for this purpose. An image can be modeled as a graph, and the segmentation problem is generally equivalent to graph problems such as finding the minimum cut, finding the minimum path and matching deformable graphs. The following four graph-based interactive image segmentation approaches: Interactive Graph Cut (IGC), Intelligent Scissors (IS), Random Walk Interactive Segmentation (RWIS) and Matching Attributed Relational Graph (MARG), will be introduced in this talk

Part 1: Text detection in compound images

Time series are an important sub-class of data in signal processing that arrange information in a time-indexed manner. They are fundamentally different from other data in this intrinsic property of order that can be and is frequently exploited. While the last few decades have generally focused on processing of single time series, the increasing volume of data streams available for many different types of applications is making it crucial that important components of underlying information are found and redundant ones are eliminated. Many new methods are being proposed in the advent of "big data" and one in particular is gaining popularity in several fields: Granger Causality. Originally proposed as an economics tool in the 1970s, it measures the extent to which one time series causes another (a statement much stronger than simple correlation). This talk will introduce several popular methods of time series analysis and discuss a few broad application areas of their use

Security is an important issue in communications and computer science. At the very beginning, Shannon started the concept of secrete-key encryption where the transmitter and legitimate receiver share a common key. Numerous researches are conducted on cryptosystems, authentication, and integrity afterwards. However, those security schemes are conventionally processed by the upper layer (such as OSI presentation Layer). Their established security are based on mainly computational complexity of attacking but not perfect secrecy. Until 1975, Wyne first proposed the idea as well as the information theoretical formulation of secrecy transmission in the physical layer. Such secrecy transmission utilizes mainly the channel characteristics instead of secrete key to achieve security. With his unique insight, a new paradigm of security researches were set up and called the physical-layer security. In recent years, signal processing techniques, such as secrecy beamforming, precoding, an

How to refer depth information from a single image automatically is a challenging topic now. Here we propose an efficient and general algorithm to infer the depth-map from a single still image using multi cues. Firstly, the foreground objects can be obtained by applying the salient map and HOS (High order statistics) computation. Then, we further apply different depth cues such as haze, focus and vanishing point to further derive a reasonable background depth model based on these depth cues. Finally, a classifier is applied here to select the best depth cue for different image categories. Furthermore, I will give a primary review on current saliency detection algorithm as it is very useful in many application scenario

Power quality issues are always big concerns in power system. The estimated loss caused by power quality issues is at least $150 billion per year. Therefore, power quality monitoring system is essential in the power system particularly in the next generation power grid, Smart Grids. In this talk, I¡¯ll first introduce the power quality issue in both legacy power grid and smart grid. Then, I¡¯ll briefly go over the single-point change point detection scheme we proposed before to remind you the fundamentals of our scheme. Finally, I¡¯ll talk about our on-going extension work on the distributed change point detection scheme.

High Content Screening (HCS) has been developed to act as an efficient tool to elucidate and define the biomolecules functions of normal and diseased cells by addressing the problem through doing large-scale detailed cellular biomedical measurement since the phenotype of cells might alter in a desired manner. In this presentation, a pipeline is discuss as to realize the investigation of drug target functioned as peroxisome proliferators in 96 sample-wells, where HCS together with image processing and machine learning is used. Non-classical approaches for illumination correction and peroxisomal structure segmentation are performed, and a method to filter images which are desirable for HCS is also proposed

Measuring a perceptual quality of an image is one of the important tasks in various applications such as image coding, processing, enhancement, and monitoring system. Although active researches

MIMO system can provide an increase in spectral efficiency which, in principal, is proportional to the minimum number of antennas in transmitter and receiver sides. With a full scattering environment assumption in which channel matrix is full rank and each sub-channel is considered i.i.d. Rayleigh fading, this enhanced spectral efficiency can be achieved. However, in case of LOS (line-of-sight) channel, LOS components dominate fully scattering components and effective channel rank is reduced to one because of the linear dependencies among LOS sub-channels. Driessen and Foschini showed in 1999 that, given properly allocated antennas at both ends, LOS channels allow spatial multiplexing and are particularly well-suited and a number of investigations have shown that high spectral efficiency and throughput can be achieved in LOS MIMO with relatively lower burden in pilot data transmission for synchronization due to low mobility. With LOS MIMO issue, non-collocated multi-antennas

Rapid development of the internet and video sharing helps to create tons of multimedia data, from which many traditionally ¡°impossible¡± problems become possible, such as machine translation, speech recognition, image annotation, and visual recognition. The problem is is how do we actually find visual information from huge amount of data when we need them? It involves two parts. One is how to efficiently represent the image or the video using some features. This is the understanding step. The other is how to index the representation to speed up the searching in real time. We need to create a clever directory structure in this step. That is the problem statement of visual information retrieval.

The demand to adapt media content (image/video) to display devices of various aspect ratios and resolutions calls for new solutions to image/video resizing. Traditional image resizing techniques are incapable of meeting this requirement since they may either discard important information (e.g., cropping) or produce distortion by over-squeezing the content (e.g., scaling).

With the number of connected computers and digital devices keeps growing, there has been a critical need for computer screen content compression and transmission technologies. The latest video coding standard HEVC has been designed mainly for camera captured video, while these computer generated video has very different characteristics compared with these camera captured video. Therefore HEVC does not fully exploit the characteri

Since 1988, Moving Picture Expert Group (MPEG) was established; Video coding has been developed for more than 20 years. In its history, PSNR plays an important role in quality assessment. However, in the latest HEVC standard meetings, subjective quality becomes more and more important for performance comparison. Video experts come up with a question -- Can PSNR really reflect to the video quality? Thus, researchers started to find methods that can evaluate human video perception, say perceptual quality assessment. Progress of perceptual quality assessment metrics (QAM) has advanced in recent years. Researchers are discovering the possibilities that how QAM can improve video coding. In this seminar, I will show our research of combing block-based QAMs with HEVC and the result of the block-based perceptual video coding algorithm for screen content

Image Understanding has been a very popular and general topic in the field of Computer Vision and Machine Intelligence in recent years. As a subtopic of this area, scene shape understanding(or background modeling) also caught many researchers' attention because of its applications to 3D world recovering from 2D images. In this Friday's seminar, I am going to introduce the field of image understanding from the prospective of scene modeling and image geometric shape understanding. Three famous groups of people and their state of the art works will be introduced to present the current development of this research area

Image segmentation is a basic processing technique in image editing and computer vision. In interactive image segmentation, the objects of interest are indicated by the interaction with human, so that the segmentation algorithms can mainly focus on object extraction rather than object identification. Graph cuts is a basic approach of interactive image segmentation. A set of graph cuts based approaches will be first introduced in this talk. Other state-of-the-art approaches like segmentation by matching attributed relational graphs (MARG) and maximal similarity based region merging (MSRM) will also be discussed

RNA-Seq is developed approach to transcriptome profiling that uses deep-sequencing technologies. Studies using this method provide us a chance of discover the complexity of eukaryotic transcriptomes and understanding their regulation rules. Furthermore, RNA-Seq also make it feasible to measure the abundance from both the gene and isoform levels with much higher accuracy than ever before. In this talk, I would introduce our own effort to develop novel statistical methodologies to dissect the human transcriptome and understand its regulation rules

The presentation include introduction of fundamental knowledge about Passive Millimeter Wave Information Detection and Imaging Technology and some of my research group¡¯s work in this field in China

The presentation include general introduction of the University of Electronic Science and Technology of China (UESTC) and the School of Electronic Engineering (SEE) of UESTC .

Cloud Computing has become a well-known primitive nowadays; many researchers and companies are embracing

With the increasing demand for rendering smooth and plausible human motion, the motion capture (or mocap) technique has been extensively used in animated movies and high-end computer games. The diverse applications of mocap data and the rapid development of motion capture systems have resulted in a large corpus of mocap data in recent years. The large amount of mocap data makes it difficult to browse raw data files, and an automated annotation technique that can classify basic motion types into multiple categories is required. Although file names and associated metadata provide a high level description of contents of mocap sequences, manual annotation of mocap sequences is labor intensive. Moreover, it is really expensive to record new data every time, so it is desirable to develop an automated technique that can classify each interval into a basic motion type, and index it for future retrieval. With such annotated mocap databases in place, a proper motion editing and author

Age related facial image processing is currently being extensively studied, and facial age group classification is one of major research topics in this area. Facial age group classification has many applications, such as age-based facial image retrieval, internet access control, security control and surveillance, biometrics, age-based human-computer interaction (HCI), age prediction systems for finding missing children, and age estimation based on the result of age group classification. One of the key ingredients in an age groups classification system is facial feature extraction. Here, two types of features, texture- and shape-based, will be discussed. Experimental results on these two types of features will be shown and compared with some state-of-the-art methods. To improve performance on the texture-based method, a feature selection technique is applied. A result on this part shows a great benefit from feature selection

In this talk, we will have a more elabroted study on objective image quality assessment (IQA) using multi-metric fusion (MMF). Extended from previous work, we develop a new fusion metric selection algorithm called the Biggest Index Ranking Difference (BIRD) that is used to select the most appropriate metric for fusion so as to reduce the complexity of the MMF method by analyzing the weight equation derived from support vector regression (SVR). Furthermore, we compare BIRD with another fusion metric selection algorithm called the Sequential Forward Metric Selection (SFMS) in terms of performance accuracy and complexity. Finally, we conduct a more thorough experimental evaluation on three publicly available image quality databases

An important step in an automatic fingerprint identification system is the process of fingerprint segmentation. The goal of segmentation is to decompose the fingerprint image into two parts: foreground (fingerprint) and background (noise). Segmentation on rolled and plain fingerprints is well-studied in literature, while latent segmentation remains to be an open research problem.

High-content screening and analysis (HCS and HCA) refer to the process of an automated acquisition of microscopy images and an automated analysis of images. The technology was developed in the late 1990s for the drug discovery market, based on the premise that cell-based multi-parametric assays were valuable early during the drug discovery process for assessing the potential of a compound to reach the market. In this talk I would introduce HCS and its application in investigation of peroxisome activity under the positive and negative treatments by different chemical compounds. Peroxisome is a ubiquitous organelle in human body which is crucial for neuronal migration, proliferation, differentiation and survival. In HCS of peroxisomal structures, image processing and machine learning approaches are designed and applied

As novel multimedia content has been created and shared in an extremely large scale everyday, how to eff ectively manage the ever growing multimedia data has become one of the most emergent problems for the multimedia research community. As a result, a large amount of attention has been drawn on the relevant topics such as large-scale image indexing and retrieval. In this talk, I will introduce the topic of near-duplicate image retrieval, and the visual bag-of-words model, which is arguably the most successful model for this task in recent years. Then, I will present my research on the improvement of visual BoW model, followed the experiments and discussion

With the growing popularity of online video sharing, energy consumption in video sharing servers has become an important issue. While saving energy in such server is achievable by utilizing low power modes, service delay has to be increased which is not desirable for video sharing services. In this paper, we present a novel energy management scheme which can jointly optimize energy and delay for the video storage system. We exploit the unique workload characteristics of video sharing services, and design a model to assist power mode selection for disks. Experiments show that our scheme achieves up to 24% more energy saving under the same delay constraint and incurs up to 57% less service delay under the same energy constraint, compared to the traditional threshold-based energy management scheme

In this research, a low©\frequency microwave link system with very high spectral efficiency is proposed. Assuming a bandwidth of 56 MHz, a data transmission rate of 2 Gbit/s shall be realized, leading to a spectral efficiency of some 40 bit/s/Hz. These extremely high values are intended to be achieved with a 2x2 MIMO system (two spatial antennas at each link end, where each antenna can be dual©\polarized). Furthermore, a transmission strategy that can switch between spatial multiplexing and diversity, thus adapting to the propagation environment is required. The goals for spectral efficiency are close to the theoretical channel capacity, and innovative signaling approaches will be necessary to realize them. Furthermore, a number of practical (implementation) non©\idealities reduce the achievable throughput, and smart, adaptive schemes are required to minimize the detrimental impact of these non©\idealities.

Security and privacy are big issues in the evolution of next generation power grid - Smart Grid. In this talk, I will introduce major security and privacy issues in Smart Grid and related research work. I'll then discuss our proposed schemes which consists of the following three parts: First, a privacy-preserving metering scheme is introduced to protect sensitive customer's power consumption data. With the scheme, we'll show that, both power consumption analysis and billing service would be accomplished without exposing customer's power consumption trace. Then, homomorphic encryption-based securing system is discussed. We first introduce a fully homomorphic encryption-based architecture for Smart Grid which is the idea case. Then, we discussed a practical partially homomorphic encryption-based solution which could be used to achieve data aggregation and statistical analysis in a privacy preserving manner. The last scheme is mainly forcusing on power quality monitoring issues

MPEG has developed a suite of international standards to support 3D services and devices, and now initiates a new phase of standardization.

Bandwidthand stabilization of network have been improved a lot over years¡¯ efforts, but when it comes to wireless transmission there is still a lot of room to improve.Video which has large size is always the biggest challenge for wireless transmission compared with other media. Video coding standard like H.264/AVC could reduce the size a lot while maintaining highquality, but at the same time it makes compressed bit stream more vulnerable to the error prone wireless channel. The other problem is that users who are receiving broadcasting video may have different resolutions. Simulcasting could solve this problem but will increase burden of network, instead spatial scalable codec could compress the video of different resolutions in a single bit stream with lower bit rate. In the talk, I will first review the basics of video coding including H.264/AVC and H.264/SVC, then I will give introduction of Multiple Description Coding (MDC) and previous works for robust spat

The drive towards decreasing upkeep costs while preserving efficient and safe operation of engines is spurring a need for intelligent failure detection and diagnosis systems. We introduce a series of comprehensive methods for the discovery, identification, and potential prediction of jet engine failures using a mixture of statistical signal processing and pattern recognition approaches. An offline detection and diagnosis method based on Mel Frequency Cepstral Coefficient (MFCC) and Code Excited Linear Prediction (CELP) features is shown to provide excellent results with vibration, but not acoustic sensor data. Results show an ability to distinguish between normal and abnormal engine operation while the engine operates in "stationary" phases, such as idling or cruising. Identification of failures in these phases of operation is also possible, although with less certainty when the sampling rate of vibration sensors approaches much lower frequencies, as would be necessitated by

The problem of depth-map inference from a single 2D video is investigated and a practical solution system is proposed in this work. We first adopt the in-focus region detection and salient map computation techniques to separate the foreground objects from the remaining background region. After that, a color-based grab-cut algorithm is used to remove the background from obtained foreground objects by modeling the background as the Gaussian mixtures. As a result, the depth map of the background can be generated by a modified vanishing point detection method while the depth values of foreground objects can be determined according to their relative sizes and occlusion information. Then, the depth map of one key frame can be propagated to the remaining frames to avoid the expensive depth-map generation for each individual 2D image

Congratulations to Martin for passing his Qual exam this morning. His PhD dissertation has the title "Robust Jet Engine Fault Detection and Diagnosis Using Vibration Sensor Data". His Qual exam committee includes: Jay Kuo (Chair), Antonio Ortega, Shri Narayanan, Keith Jenkins and Alexandra Tartakovsky (Outside Member)

Congratulations to Joyce for passing her defense this afternoon. Her PhD dissertation title is "Efficient Methods for Enhancing Secure Network Codes," and her dissertation committee includes: Jay Kuo (Chair), Zhen Zhang and Leana Golubchik (Outside Member)

Audio signal processing is similar to video/image processing in many ways, and is different too, in many ways. In tomorrow's talk, such parallels would be drawn to facilitate leveraging of the image processing tools/frameworks for audio processing keeping in mind applications like filtering, signal decomposition and representation, signal analysis, etc. In particular, signal representation approaches via sparse representation would be discussed. Finally, a more complex problem of audio scene recognition would be introduced and discussed

Visual analysis of three-dimensional (3D) Coronary Computed Tomography Angiography (CCTA) remains challenging due to large number of image slices and tortuous character of the vessels. We aimed to develop an accurate and automated algorithm for detection of coronary artery lesions compared to human expert interpretation. Our knowledge-based automated algorithm consists of centerline extraction which also classifies 3 main coronary arteries and small branches in each main coronary artery, vessel linearization, lumen segmentation with scan-specific lumen attenuation ranges, and lesion location detection. Presence and location of lesions are identified using a multi-pass algorithm which considers expected or "normal" vessel tapering and luminal stenosis from the segmented vessel. Expected luminal diameter is derived from the scan by automated piecewise least squares line fitting over proximal and mid segments (67%) of the coronary artery, considering small branch locations. We

The problem of depth-map inference from a single 2D video is investigated and a practical solution system is proposed in this work. First, foreground objects are coarsely identified based on techniques of in-focus region detection and salient map computation. Then, we develop a modified vanish point detection method to generate the depth-map for the background, and assign depth values to those segmented foreground objects based on their relative sizes and occluding relationship. Finally, we propagate the depth map sequence using motion vector estimation

Engine monitoring systems largely rely on periodic examination of pressure and temperature during manual maintenance. In our recent work, we devised method of monitoring and detecting problems in the mechanical operation of such systems using vibration sensor readings. All past work was done using data from laboratory experimental settings, where characteristics of normal and abnormal engine operations were known. Our current work is focused on the difficult, yet real-world scenario, of detecting problems and predicting breakdowns when only normal operating data is available. In this seminar, I will introduce the abstract problem of single-class anomaly detection, demonstrate common approaches, and introduce a case study in engine failure prediction

Traditional image steganalysis techniques are conducted with respect to the entire image. In this work, we aim to differentiate a stego image from its cover image based on steganalysis results of decomposed image blocks. As a natural image often consists of heterogeneous regions, its decomposition will lead to smaller image blocks, each of which is more homogeneous. We classify these image blocks into multiple classes and find a classifier for each class to decide whether a block is from a cover or stego image. Then, the steganalysis of the whole image is conducted by fusing decision results of all image blocks.

A new joint estimation of timing, carrier frequency offset and channel for millimeter wave transmission with unsynchronized clocks is proposed. Practical wireless communication achieving very high throughput has many limitation on its implementation so many proposed algorithms have been considered techniques which can be existed only on research papers. However, millimeter wave transmission for wireless backbone has been focused in countries having huge territory and low mobility characteristic of this application can give a small clue for algorithms to accomplish high throughput practically. In this presentation, 2x2 MIMO and polarization are combined to maximized channel throughput and system clocks for each Tx and Rx antenna system are not synchronized so, from each Rx antenna system¡¯s point of view, it suffers 4 possible offsets in timing frequency offset (TFO) and carrier frequency offset (CFO) and joint estimation for these 2 TFO, 2 CFO and channel plays a key role bec

Machine learning technologies have been widely used in many applications, including data information processing and visual information processing. Recently, due to huge amount of text and audio data available in digital form, we observed some successful examples in learning-based text classification and speech recognition. In these days, a lot of images and videos are available over the internet. Some machine learning based methods have been applied to deal with image/video processing problems. Here I will give two examples, image retrieval with relevance feedback and image quality assessment, to show learning-based methods in visual information processing

In this talk, I will present a new methodology for objective image quality assessment with multi-metric fusion (MMF). First, I will review some previous work. The current research is motivated by the observation that there is no single metric that gives the best performance scores in all situations. To achieve MMF, we adopt a regression approach. It is shown by experimental results that the proposed MMF metric outperforms all existing metrics by a significant margin

Content-Centric Networking (CCN) is a new network architecture (and communication paradigm) which has been proposed to replace the current IP-based Internet. It can be seen as a generalization of IP by addressing content instead of end-points. In other words, the network allows the user to pull content by name instead of some physical address. CCN achieves this location independence by utilizing in-network caching and content-name-based routing.

Segmentation is an important step in automatic fingerprint identification systems. While tremendous progress has

The development of the second generation sequencing technology has profoundly changed the approach to metagenomics through massive and parallel, but low cost, sequencing of microbial species unsuitable for laboratory cultivation. One challenge in many projects in metagenomics is to develop a robust de novo sequence

Secrecy and robustness are both vitally relevant to the design of communication networks, but they are often treated independently in network coding (NC) literature. While a wireless network can leverage network coding to improve the robustness, the redundancy added by broadcast accentuates the vulnerability to arbitrarily located eavesdroppers. We first quantify the tradeoff between secrecy and robustness in secret sharing. We then apply constraints for secrecy and robustness specified by network operators to design an algorithm that minimizes the number of key exchanges. The proposed algorithm incorporates a windowed coding method with feedback to deal with bursty erasures and allows nodes to recover from erasures over a period of time. Moreover, this method increases secrecy by reducing the benefit an eavesdropper may derive from relay-injected redundancy. Compared with the existing schemes on secure network coding, our algorithm does not restrict the number of edges an

Video copy detection is pretty hot these days, and there are 22 teams participated in the CCD sub-task of TrecVid. It is very important for the topics like Copyright control, Business intelligence and advertisement tracking, Law enforcement investigations.

Visual analysis of three-dimensional (3D) Coronary Computed Tomography Angiography (CCTA) remains challenging due to large number of image slices and tortuous character of the vessels. We aimed to develop an accurate, automated algorithm for detection of significant and subtle coronary artery lesions compared to expert interpretation. Our knowledge-based automated algorithm consists of centerline extraction which also classifies 3 main coronary arteries and small branches in each main coronary artery, vessel linearization, lumen segmentation with scan-specific lumen attenuation ranges, and lesion location detection. Presence and location of lesions are identified using a multi-pass algorithm which considers expected or "normal" vessel tapering and luminal stenosis from the segmented vessel. We applied this algorithm to 21 CCTA patient datasets, acquired with dual-source CT, where 7 datasets had 17 lesions with stenosis greater than or equal to 25%. Our algorithm identified 1

Video coding has been an widely adopted application for decades. The contemporary video coding standard, H.264, was originally designed for low bit-rate web-casting environment. For the emerging demand of high definition and high quality video transmission such as HDTV, H.264 fails to achieve a satisfiable compression performance. In this talk, two major successor work of H.264 will be investigated: High Efficiency Video Coding (HEVC) and Multi Order Residual Coding (MOR). HEVC, also known as H.265, is an ongoing standard jointly developed by Joint Collaborative Team on Video Coding (JCT-VC). HEVC is architecturally similar with H.264, while a lot of new improvements in each module of the coding process. Multi Order Residual Coding, on the other hand, is architectural innovation based on H.264, where various of coding modules (like prediction, DCT, quantization, etc) are re-designed and cascaded on the existing coding architecture. The two future coding schemes will be individually analyzed in this talk

Computer-aided segmentation of the boundaries of the chambers of heart and the whole heart from images obtained by various imaging modalities such as Magnetic Resonance Imaging (MRI), Computed Tomography (CT), Positron Emission Tomography (PET), Single Photon Emission Computed Tomography (SPECT) or Ultrasound Imaging (US) is a prerequisite for the derivation of several quantitative cardiac parameters including functions and perfusions which have great clinical importance. Over the last several decades, numerous computerized methods have been developed to tackle this problem. Seminal works were heuristic, but recent studies employ sophisticated techniques such as partial differential equations, differential geometry, statistical inference and machine learning. Recent work exploits available cues from cardiac anatomy such as geometry and visual appearance and prior knowledge from statistical modeling. In addition, new minimization and computational methods have been adopted with improved computational speed and robustness. In this talk, I will summarize techniques used for automatic segmentation of chambers of the heart and the whole heart from images obtained by MRI, CT, Nuclear Imaging (SPECT and PET) and US

Although timing recovery is the most fundamental part of synchronization process in wireless communication and many people know theoretical knowledge of it, how to implement it efficiently is not well known to students. Digital interpolator using farrow structure is the practical algorithm still mostly used in almost modern wireless transceiver chips even though the research paper regarding it was published about 10 years ago. In this talk, I will first give the very basics and theoretical algorithms of digital timing recovery, and then I will concentrate on how farrow structure works in real transceiver design and why it is more effective than other algorithms by comparison

Recent years have witnessed an explosive growth in video streaming and sharing services. To support these large-scale Internet services, data centers need to be expanded in both scale and density. Their energy efficiency has become a central issue of both economic importance and environmental urgency. The emerging cloud computing paradigm has promised high availability and instant scalability for Internet services. Cloud-based video delivery requires multi-layer, distributed and virtualized systems, which make energy efficiency a more challenging problem. In this talk, I will first give an overview of the problem and discuss some energy issues in video data centers. In particular, I will introduce an approach to profiling and modeling the energy consumption in video services. Then, I will zoom into the storage sub-system, and focus on caching and scheduling design for media servers. I will introduce a novel algorithm that achieves energy-aware caching and scheduling. Essentially, energy and performance metrics can be modeled and used in cache replacement and scheduling decisions. In this way, an energy-performance trade-off can be applied and optimized

The deep sequencing of second generation sequencing technology has enabled us to study complex biological structures having multiple DNA units such as transcriptomics, metagenomics, and so on. Unlike general genome sequence assembly, a DNA unit of these biological structures may have multiple copies with small or big structural variations and/or SNPs simultaneously in one sample. Therefore, the deep sequencing, which is hard to achieve with expensive Sanger sequencers, is required to figure out such variations concurrently. This proposal focuses on de novo transcriptome assembly which requires simultaneous assembly of multiple alternatively spliced gene transcripts. In practice, the de novo transcriptome assembly is the only option for studying the transcriptome of organisms that do not have reference genome sequences, and it can also be applied to identify novel transcripts and structural variations in the gene regions of model organisms. We propose a de novo transcriptome assembly called WEAV which first partitions RNA-seq reads into clusters, and then for each cluster, applies a de Bruijn graph-based method to simultaneously identify multiple alternatively spliced gene transcripts. WEAV provides a novel approach to error-correction by using a fast mapping program called PerM, and a partition-ligation algorithm for transcript assembly. We used WEAV to assemble 22.68 million (after pre-processing) 100bp human RNA-seq reads into 156,494 contigs that were longer than 200bp. 96.3% (specificity) of these contigs were mapped onto either RefSeq, Gencode or human Genome sequences (hg19), and they covered >72% sequenced bases annotated in RefSeq and Gencode. These high sensitivity and specificity showed the exceptional power of WEAV for transcriptome assembly

The title of Sphinx' thesis proposal is "Vibration and Audio Signal Processing Using CELP and HHT Features". The Guidance Committee includes Jay Kuo (Chair), Shri Narayanan, Jerry Mendel, Keith Jenkins and Tu-Nan Chang (Outside Member)

An adequate feature set plays a key role in many signal classification and recognition problems. This is a challenging problem due to the nonlinearity and nonstationary characteristics of real world signals, such as engine acoustic/vibration data, environmental sounds, speech signals and music instrument sounds. Some of traditional features such as the Mel Frequency Cepstral Coefficients (MFCC) may not offer good performance. Other features such as those based on the Matching Pursuit (MP) decomposition may perform better, yet their complexity is very high. In this research, we consider a new feature set that can be easily generated in the model-based signal compression process, known as the Code Excited Linear Prediction (CELP) features. The CELP-based coding algorithm and its variants have been widely used to encode speech and low-bit-rate audio signals. In this research, we examine two applications based on CELP-based features

The title of her thesis proposal is "Efficient Techniques for Enhancing Secure Network Codes,' and her Guidance Committee includes Jay Kuo (Chair), Jerry Mendel, Andy Molisch, Zhen Zhang and Ming-Deh Huang (Outside Member)

Network coding has been studied extensively in the last decade, during which it has been linked to applications in throughput gain, error correction, robustness to non-ergodic link failures, confidentiality, security against Byzantine attacks, P2P, storage distribution, etc. It can be efficiently implemented through low complexity linear operations, over both wired and wireless networks. Network coding may also be utilized in both centralized and decentralized designs, where it has been shown to outperform traditional routing techniques. However, many of the studies into network coding applications were conducted in isolation, and therefore do not consider the holistic implementation with network requirements on security, efficiency, and robustness against erasures. Joint examination yields design considerations that require flexible parameters to analyze tradeoffs between desired algorithmic features. For the first topic, a novel algorithm that achieves robust and secure sharing jointly based on network coding among multiple trusted peers in wireless erasure networks is investigated in this work. In the considered communication model, an eavesdropper can take advantage of the broadcast medium to tap messages along a min-cut. The fundamental tradeoff between secrecy, robustness and efficiency enforced by the wireless environment is examined. In situations where there is no secure capacity, we give a convolutional NC scheme that achieves weak secrecy to decrease the number of symmetric keys required. We further propose methods to increase communication efficiency using the algorithm at the cost of robustness or privacy. Finally, we show that network erasures can actually increase the amount of secrecy in the proposed scheme but at the cost of decreased efficiency. For the second topic, an algorithm is presented that reduces the overhead communication cost traditionally incurred by all random linear network coding. Unlike the majority of secure network coding algorithms, the proposed algorithm continues a very recent trend that instead focuses on preventing the adversary from ascertaining any global or local mappings. Without the global coding matrix, the adversary is unable to recover the secret source message, hence the adversary¡¯s primary concern is how to recover the global coding matrix given its wiretap edge observations. Our algorithm is novel in the sense that intermediate nodes code messages based on the secretly indexed set of local coding vectors. By reducing the size of this finite pool, we may achieve enormous gains in overhead savings, which increase utilizable network capacity. The impact of the field size and the size of the vector pool on both security and efficiency is discussed

CCongratulations to Jing Zhang for passing her Qual exam today (1/12/2011). Jing's proposal title is "Road to Decipher the Splicing Code." Her guidance committee includes: C.-C. Jay Kuo (Chair), Liang Chen (Co-Chair), Jerry Mendel, Richard Leahy, Yan Liu and Fengzhu Sun (Outside Member)

Congratulations to Sungje Cho for passing his Qual exam today (1/12/2011). Sungje's proposal title is "Techniques for De Novo Sequence Assembly: Algorithms and Experimental Results." His guidance committee includes: C.-C. Jay Kuo (Chair), Ting Chen (Co-Chair), Antonio Ortega, Richard Leahy and Andrew Smith (Outside Member)

Congratulations to Selina Chu for passing her defense today. Her thesis title is "Recognition and Characterization of Unstructured Environmental Audio". Her thesis committee includes: Shri Narayanan (Chair), Jay Kuo (co-Chair), Cyrus Shahabi and Keith Jenkins (Outside Member)

Congratulations to Steve Cho for passing his Qual. exam today. His thesis title is "Block-based image steganalysis: algorithm and performance evaluation". His Qual exam committee includes: Jay Kuo (Chair), Antonio Ortega, Shri Narayanan, Keith Jenkins and Ming-Deh Huang (Outside Member)

Image steganography embeds secret messages so that no one except the intended recipients can detect the presence of secret messages. In contrast, image steganalysis tries to detect the presence of secret messages hidden in images and eventually extract the secret messages.

General purposed computing on graphics processing units (GPGPU) is the technique of using a GPU, which typically handles computation only for computer graphics, to perform computation in applications traditionally handled by the CPU. It is made possible by the addition of programmable stages and higher precision arithmetic to the rendering pipelines, which allows software developers to use stream processing on non-graphics data. Recently there is trend shifting GPUcomputing on the mobile devices. As a combination of the fast computing ability of GPU and the portability convenience of mobile devices, numerous of novel applications can be realized. In this talk, we will focus on its application of augmented reality (AR). The current technology trend and research works of mobile AR will be discussed. The future vision of GPGPU and AR will also be discussed in the talk

A first-order-residual/second-order-residual (FOR/SOR) based video coding algorithm that incorporates the super macroblock (SMB) and the inter-frame stripe prediction (ISP) technique is proposed for high definition (D) video coding in this work. We first examine the limitation of the SMB in high-bit-rate coding, and show that a simple extension of the block-size is not sufficient to get a significant improvement in coding performance. Then, we introduce the FOR/SOR coding technique to resolve this issue. For the for coding, we apply a larger quantization stepize and motion-predictive coding to remove the temporal correlation. Then, for the SOR coding, we propose an efficient prediction method, called the inter-frame stripe prediction (ISP) technique, to remove the structured residuals after the FOR coding to achieve a higher coding gain. It is demonstrated by experimental results that the proposed FOR/SOR algorithm outperforms H.264/AVC by a significant margin; namely, about 21.57 percent in the bit rate saving or 1.09dB in the PSNR improvement

Automatic Segmentation of multimedia content is still a relatively open field, owing to the difficulty of separating of real-world signals into their semantic components, a feat easily accomplished by the human brain. Advances in speech and music understanding have made segmentation useful, although imperfect, for these specialized fields, but a solution to the general case is still out of reach. The first hurdle along the path to scene separation and analysis, is silence detection - a surprisingly difficult and elusive task. With practical applications ranging from entertainment to security, a robust automatic audio segmentation system would also greatly improve the results of any processing methods that depend on pre-segmented audio

Super Resolution (SR) image reconstruction finds applications not only to overcome the inherent resolution limitations of low-cost imaging sensor, but also to cope up with the capabilities of high-resolution displays like HD display units. Imagine having an algorithm which would perform real time SR and allow you to watch your entire old personal video library collection on HDTV! Or think of the possibility where you could capture images from your cell phones and still have a high resolution photograph. Digital media revolution provides enough motivation to pursue a good SR algorithm. SR from single low resolution images is an ill-posed problem as compared to the SR from multiple row resolution images. Observed low resolution image can be considered as sub-sampled and blurred version of the expected high resolution counterpart. The fundamental reconstruction constraint for SR is that the recovered image should reproduce the observed low resolution image. Given a low resolution image as constraint and the only ground truth, solution from the reconstruction constraint is not unique. There are several SR algorithms based on signal processing tools in both spatial and frequency domains, based on statistical properties, and many more ¨C and all of these have limited modeling capabilities, limited domain sets or limited range sets. With success of Machine Learning tools in several fields of vision, it is not surprising that one would expect effective and robust algorithm using ML techniques. In this presentation, I will give state the problem SR problem, give an overview of existing algorithms and present a frame-work based on ML techniques

Embedding pico/femto base-stations and relay nodes in a macro-cellular network is a promising method for achieving substantial gains in coverage and capacity compared to macro-only networks. These new types of base-stations can operate on the same wireless channel as the macro-cellular network, providing higher spatial reuse via cell splitting. However, these base-stations are deployed in an unplanned manner, can have very different transmit powers, and may not have traffic aggregation among many users. This could potentially result in much higher interference magnitude and variability. Hence, such deployments require the use of innovative cell association and inter-cell interference coordination techniques in order to realize the promised capacity and coverage gains. In this presentation, I introduce conventional interference management algorithm used in macro cell planning and new paradigms for design and operation of such heterogeneous cellular networks. Specifically, I focus on cell splitting, range expansion, semi-static resource negotiation on third-party backhaul connections, and fast dynamic interference management for QoS via over-the-air signaling

Distributed wireless mesh networks have gained significance in their ability to solve the last-mile problem in information distribution. They are ideal in situations that require temporary, fast ,and inexpensive deployment, such as in social, emergency, and battlefield environments. Because of the nature of wireless multihop, these networks suffer from interference as well as physical security vulnerabilities. This talk will feature a novel algorithm that achieves robust and secure secret sharing jointly based on network coding (NC) among multiple trusted peers in wireless erasure networks. In the considered communication model, an eavesdropper can take advantage of the broadcast medium to tap messages along a min-cut. The fundamental tradeoff between secrecy, robustness and efficiency enforced by the wireless environment is examined. In situations where there is no secure capacity, a convolutional NC scheme will be proposed that achieves weak secrecy to decrease the number of symmetric keys required, including several methods to increase the efficiency at the cost of robustness or privacy. At the conclusion, possible future approaches to NC-based secrecy in wireless networks will be introduced

We live in a "Multimedia age" where millions of videos and images are seemlessly produced, stored and shared every day. Online video sharing websites like YouTube, Google Videos etc., have billions of videos which need efficient indexing schemes for fast search and retrieval of query videos. These indexing schemes should not only be easier and compact to implement but should also be faster and accurate. To address the problem of efficiently indexing large databases, researchers have recently introduced new hashing techniques such as Spectral Hashing, SPEC-hashing, Semantic Hashing using RBMs,etc. The goal of this talk is to discuss the challenges in large video databases, and introduce new approaches for video representation and present our Two-stage Semantic-based Hashing technique to index large video datasets. In this talk, several topics such as Deep Belief Nets, Semantic Similarity Distance Learning and Video Representation will be discussed

Early detection and accurate assessment of heart disease is crucial in the identification of patients risk for the highly common yet potentially preventable events of myocardial infarction and sudden cardiac death (SCD). In this talk, the talk consists of 2 main stories. First one is 'Heart Segmentation (a Review). Computer-aided segmentation of the boundaries of the left ventricle (LV) and the whole heart from images of a beating heart obtained by various imaging modalities such as Magnetic Resonance Imaging (MRI), Computed Tomography (CT), Positron Emission Tomography (PET), Single Photon Emission Computed Tomography (SPECT) or Ultrasound Imaging (US) is a prerequisite for the derivation of several quantitative cardiac parameters including functions and perfusions which have great clinical importance. I will talk about the numerous computerized methods that have been developed to tackle this problem. The second one is 'Lesion Detection in coronary arteries from CCTA', which I am co-researching with Cedars-Sinai Medical Center's research team. Current CCTA image analysis techniques for detection and characterization of coronary plaques are manual, time-consuming and subjective due to high observer variability. The research project aims to refine, validate, and assess prognostic value of a fully automatic software system. I will present the current stage on the automatic lesion detection including centerline extraction, linearization of arteries, and lesion detection

An efficient and robust image matching algorithm using a set of selected SIFT descriptors is investigated in this work. The proposed selected SIFT can offer more robust and stable image matching results.

His thesis title is Physical Layer Multicasting with Opportunistic Multicast Scheduling, and his thesis committee includes Jay Kuo(chair), Jerry Mendel, Andreas F. Molisch, Alexandros G. Dimakis and Wlodek Proskurowski(Outside Member)

His thesis title is Low Complexity and Highly Efficient Prediction Techniques for Video Coding, and his thesis committee includes Jay Kuo(chair), Richard Leahy, Jerry Mendel, Sandeep K. Gupta and Aichiro Nakano(Outside Member)

Her thesis title is Mocap Data Compression: Algorithms and Performance Evaluation, and her thesis committee includes Jay Kuo(chair), Shri Narayanan and Aichiro Nakano(Outside Member)

Abstract: Motion capture (Mocap) data stores motion data in 3D. In order to store, retrieve, and transfer Mocap data efficiently. Two different compression frameworks designed to optimize different objective function for different application purposes will be discussed and presented in this talk. The first one is based on my previous work aiming at preserving nearly lossless content. With the new temporal and spatial information handling, we can now achieve 50:1 compression, which is 2.5 times saving than the previous work. The second one is perception based compression, which aims for further compression with visually pleasant quality. We achieve on average 160:1 compression ratio in this framework

Abstract

Abstract:

Abstract:

Abstract:

Abstract:

Abstract:

Abstract:

Congratulations to Naco for passing her Qualifying exam! Her thesis proposal is entitled with "Design of Feature-Preserving Thumbnails for 3D Mesh Database Management". Her thesis guidance committee includes: Jay Kuo (Chair), Gerard Medioni, Aiichiro Nakano, Suya You and Tunan Chang (Outside Member)

Abstract:

4/24-4/30 15 Roy Lou

Abstract:

Congratulations to Jonghye Woo for winning the 2009-2010 Best Disseration Award - Best Experimental Work

Congratulations to Shahryar Karimi-Ashtiani for passing his PhD. defense today. His thesis title is "Theory and Simulation of Diffusion Magnetic Resonance Imaging on Brain's White Matter". His disseration committee includes: Jay Kuo (Chair), Richard Leahy and Minbir Singh

Abstract:

Abstract:

Abstract:

Abstract:

Congratulations to Youngmin Kwak for passing his defense today (3/23). His thesis title is "Advanced Liquid Simulation Techniques for Computer Graphis Applications". His disseration committee includes: Jay Kuo (Chairman), Jerry Mendel and Aichiro Nakano (Outside Member)

Congratulations to Qi Zhang for passing his defense this morning. His dissertation title is "Advanced Techniques for High Fidelity Video Coding" and his disseration committee includes: Jay Kuo (Chairman), Antonio Ortega and Cyrus Shahabi (Outside Member)

Congratulations to Angel for passing her defense today. Her thesis title is "Advanced Intra Prediction Techniques for Image/Video Coding". Her disseration committee includes: Jay Kuo (Chair), Antonio Ortega and Tunan Chang

Abstract:

Abstract:

Abstract:

Abstract:

Congratulations to May-chen for passing her Qual exam today. Her proposal title is "Mocap Data Compression for Human Characters Animation". Her thesis Guidance Committee includes: Jay Kuo (Chair), Antonio Ortega, Shri Narayanan, Suya You and Aiichiro Nakano (Outside Member)