Next Scheduled Talk
Speaker Ken Sooknanan
Title Content Analysis of Underwater Video Sequences for Automatic Seabed Change Detection
Time & Venue Printing House Hall - 11:30am 2
nd June
Abstract Underwater surveys normally involve the recording of long videos of the ocean bed. These recordings are then tediously analysed by a marine biologist to detect specific items of interest e.g. changes of the ocean bed etc. This project aims to simplify the analysing phase of these surveys, by automatically detecting the critical portions of the video when (if any) major changes in the seabed had occurred, and then summarizing these findings in short sequences to the user. In this talk, some of the problems encountered in analysing underwater video sequences are presented, along with some of the possible solutions investigated. Previous work involving content analysis specific to this area of underwater surveillance videos would also be discussed.
Speaker Finnian Kelly
Title Speaker Verification for Biometrics
Time & Venue Printing House Hall - 11:30am 2
nd June
Abstract Biometrics involves the use of intrinsic physical or behavioural traits of humans to verify their identity. Speaker verification offers a less invasive alternative to iris and fingerprint authentication.
Typical applications involve banking by phone and access authentication. Key challenges include dealing with mismatched conditions and natural changes in the speaker's voice due to illness or ageing. This talk outlines the implementation of a Gaussian Mixture Model based speaker verification system. Different training strategies, including a split and merge approach are described. Results are presented on the YOHO Speaker Verification database. Normalization methods to deal with mismatched conditions are outlined. Some 'higher level' voice features and their potential to provide robustness to ageing are introduced.
Upcomming Talks
Previous Talks
26th May 2010
Speaker Marcin Gorzel
Title VIRTUAL ACOUSTIC RECORDING: AN INTERACTIVE APPROACH
Abstract/Details
Virtual acoustic recording refers to the capture of real-world acoustic performances in reverberant spaces and their subsequent plausible reproduction in a virtual version of the original performance space, otherwise known as a Virtual Auditory Environment (VAE). An important aspect in the quest for realism in such auditory scene synthesis is user interaction. That is, how the movements of a person listening to the virtual auditory scene directly influences the scene presentation. Such 'walkthrough auralization presents several challenges for production engineers, the most significant of which is the generation of the correct room acoustic response due to a given source-listener position. In particular, the correct direction of arrival of the direct sound and early reflections must be maintained since these signals contain the most vital cues for localization of acoustic sources.
Speaker Róisín Rowley-Brooke
Title The Digital Restoration of Degraded Historical Manuscripts
Abstract/Details
Until fairly recently the only way of improving the readability of degraded manuscripts was through manual, and often destructive methods. The advancement of image processing has allowed for manuscript restoration techniques that avoid the possibility of doing further damage to document. The aim of this talk is to outline briefly the main types of degradation that can occur such as bleed through, fading and erosion of the writing medium due to poor storage. Methods for manuscript improvement will be discussed; the use of adaptive histogram equalization for enhancement of faded text, two approaches to modelling palimpsest (or bleed-through) documents, and also a method for recto/verso registration - an important pre-processing step in bleed-through removal.
19th May 2010
Speaker Dr. David Corrigan
Title Non-parametric Texture Synthesis in the Wavelet Domain
Abstract/Details
The goal Texture Synthesis is to generate new and typically larger textures that are perceptually similar to an example texture. It is an important concept in the domain of cinema-post production as it can be used to generate textures to map onto the surfaces of computer generated objects. This talk will outline the techniques we have developed for texture synthesis that operate wholly or partly in the wavelet domain. The basis of each technique is a non-parametric method which
synthesises wavelet coefficients at the coarsest level of the wavelet transform. For textures with small
features, it is shown that the coarse resolution search is sufficient for realistic synthesis. Two alternative
approaches are proposed for more structured synthesis. The first is a non-parametric refined multiscale
synthesis which synthesises coefficients at all levels of the wavelet tree. The second uses the coarse
resolution search to facilitate a patch-based synthesis. These approaches are shown to synthesise realistic
textures at a much reduced computational expense than other non-parametric techniques. The fidelity of
the synthesised texture is also more robust to variations in texture scale.
Slides David's Slides
18th May 2010
Speaker Prof. Barak Pearlmutter
Title What Sparsity Means to Me.
Abstract/Details
Our computational toolbox holds many highly efficient methods, like least-squares linear fits. Our theoretical cookbook lists many intractable recipes, like calculating and marginalizing over posteriors using complete detailed models and accurate priors. Sparseness supplies a surprising bridge between these: methods to find "sparse" solutions to under-constrained problems can be surprisingly
efficient, and distributions based on a simple sparsity criterion are surprisingly effective at encoding prior structure. In this talk we will explore a variety of ways in which sparseness can be represented mathematically and exploited computationally.
12th May 2010
Speaker Eric Risser
Title Texture Synthesis
Abstract/Details
The goal Texture Synthesis is to generate new and typically larger textures that are perceptually similar to an example texture. It is an important concept in the domain of cinema-post production as it can be used to generate textures to map onto the surfaces of computer generated objects. This talk will outline the techniques we have developed for texture synthesis that operate wholly or partly in the wavelet domain. The basis of each technique is a non-parametric method which
synthesises wavelet coefficients at the coarsest level of the wavelet transform. For textures with small
features, it is shown that the coarse resolution search is sufficient for realistic synthesis. Two alternative
approaches are proposed for more structured synthesis. The first is a non-parametric refined multiscale
synthesis which synthesises coefficients at all levels of the wavelet tree. The second uses the coarse
resolution search to facilitate a patch-based synthesis. These approaches are shown to synthesise realistic
textures at a much reduced computational expense than other non-parametric techniques. The fidelity of
the synthesised texture is also more robust to variations in texture scale.
14th April 2010
Speaker Francois Pitie and Dan Ring
Title Nuke and Plugin Development
Abstract/Details This talk will outline some of the recent work in plugin development for the Nuke platform.
31st March 2010
Speaker Gary Baugh
Title Semi-automatic Motion Based Segmentation using Long Term Motion Trajectories
Abstract Semi-automated object segmentation is an important step in the cinema post-production workflow. We propose a dense motion based segmentation process that employs sparse feature based trajectories estimated across a long sequence of frames, articulated with a Bayesian framework. The algorithm first classifies the sparse trajectories into sparsely defined objects. Then the sparse object trajectories together with motion model side information are used to generate a dense object segmentation of each video frame. Unlike previous work, we do not use the sparse trajectories only to propose motion models, but instead use their position and motion throughout the sequence as part of the classification of pixels in the second step. Furthermore, we introduce novel colour and motion priors that employ the sparse trajectories to make explicit the spatiotemporal smoothness constraints important for long term motion segmentation.
24th March 2010
Speaker Prof. Anil Kokaram
Details Anil gave an account about his recent visit to the west coast of the USA and how he talks about sigmedia when on tour.
4th March 2010
Speaker Andrew Hines
Title Measuring Sensorineural Hearing Loss with an Auditory Peripheral Model
Abstract Hearing loss research has traditionally been based on perceptual criteria, speech intelligibility and threshold levels. The development of computational models of the auditory-periphery has allowed experimentation via simulation to provide quantitative, repeatable results at a more granular level than would be practical with clinical research on human subjects. Model outputs can be assessed by examination of the spectro-temporal output visualised as neurograms. The effect of sensorineural hearing loss (SNHL) on phonemic structure was evaluated using two types of neurograms. A new systematic way of assessing phonemic degradation is proposed using the outputs of an auditory nerve model for a range of SNHLs.
10th February 2010
Speaker Dr. JiWon Yoon
Title Bayesian Inference for Single Molecule Fluorescence Microscopic image processing
Abstract Using fluorescence microscopy with single-molecule sensitivity, it is now possible to follow to movement of individual fluoro-phore tagged molecules such as proteins and lipids in the cell membrane with nano-meter precision. Diffusion or directed motion of molecules on the cell can be investigated to elucidate the structure of the cell membrane by tracking the single molecules. There are mainly three steps in processing data and tracking the molecules from the sequential images: filtering (de-noising), spot detection and tracking. In this talk, we will give a presentation on both filtering and tracking techniques.
First of all, we have recently developed a robust de-noising algorithm in a Gibbs scheme. This algorithm embeds Gaussian Markov Random Field (GMRF) prior to explain the properties of the images. Since this algorithm is based on Bayesian framework, we do have few systematic parameters to be tuned. The performance of this algorithm is compared with several conventional approaches including Gaussian filter, Weiner filter and Wavelet filter.
We also developed several multi-target tracking algorithms in a Bayesian framework. Roughly, we will retrieve the concept of single target tracking and multi-target tracking. Then, the marginalized Markov Chain Monte Carlo Data Association (MCMCDA) which is originally proposed by Oh is presented. Marginalized MCMCDA is a fully off-line system and it infers most systematic parameters which are commonly fixed in tracking society.
Bio He received the B.Sc. degree in information engineering at the SungKyunKwan University, Korea. He obtained the M.Sc. degree in School of informatics at the University of Edinburgh UK in 2004 and the Ph.D. degree in signal processing group at the University of Cambridge UK in 2008 respectively. In 2008, he moved to department of Engineering science, the University of Oxford, UK to do postdoctoral research. He is currently a Research Fellow with Statistics department, Trinity College Dublin, Ireland. His research interests include Bayesian statistics, Machine Learning, data mining, Network Security and Biomedical engineering. He has worked on applications in brain signals, cosmology, biophysics and multimedia.
4th February 2010
Speaker Dr. Sid-Ahmed Berrani, Orange Telecom
Title TV Broadcast Analysis and Structuring: Advances and Challenges
Abstract In order to make use of the large number of TV broadcasts through novel services like Catch-up TV or TV-on-Demand, TV streams have to be precisely and automatically segmented, annotated and structured. The exact start time and the exact end time of each of the broadcasted programs have to be determined. Each extracted program has then to be classified, annotated and indexed.
The aim of this talk is to provide an overview of novel TV services and to highlight the need for powerful audio-visual content-based analysis techniques in order to build these services. Technical constraints will be also discussed. Our work toward building a fully automatic system for TV broadcast structuring will be then described. Finally, open issues and challenges will be presented.
Bio Sid-Ahmed Berrani received his Ph.D. in Computer Science in February 2004 from the University of Rennes 1, France. His Ph.D. work was carried out at INRIA, Rennes and was funded by Thomson R&D France. It was dedicated to similarity searches in very large image databases. The Ph.D. thesis of Sid-Ahmed Berrani received the SPECIF Award from the French Society of Education and Research in Computer Science. He then spent 6 months as a Research Fellow in the Sigmedia Group at the University of Dublin, Trinity College, where he worked on video indexing. Since November 2004, Sid-Ahmed Berrani has been a researcher at Orange Labs - France Telecom in Rennes, France. He is currently leading R&D activities on video indexing and analysis for media search services. In particular, he has focused on video analysis techniques for TV broadcast structuring and video fingerprinting.
15th January 2010
Speaker Prof. Nick Kingsbury, Cambridge University
Title Iterative Methods for 3-D Deconvolution with Overcomplete Transforms, such as Dual-Tree Complex Wavelets.
Abstract Overcomplete transforms, such as complex wavelets, can offer more flexible signal representations than critically-sampled transforms. They have been shown to perform well in image denoising benchmarks, and we have therefore been developing iterative wavelet-based regularisation algorithms for more demanding applications such as image and 3D-data deconvolution. In this talk we will briefly describe the characteristics of complex wavelets that make them well-suited to such tasks, and then we will describe an algorithm for wavelet-based 3-dimensional image deconvolution which employs subband-dependent minimization and the dual-tree wavelet transform in an iterative Bayesian framework. This algorithm employs a prior based on an extended Gaussian Scale Mixtures (GSM) model that approximates an L0-norm, instead of the conventional L1-norm, to provide a sparseness constraint in the wavelet domain. Hence it introduces spatially varying inter-scale information into the deconvolution process and thus achieves improved deconvolution results and faster convergence.
Bio Nick Kingsbury is Professor of Signal Processing at the University of Cambridge, Department of Engineering. He has worked in the areas of digital communications, audio analysis and coding, and image processing. He has developed the dual-tree complex wavelet transform and is especially interested in the application of complex wavelets and related multiscale and multiresolution methods to the analysis of images and 3-D datasets.
16th December 2009
Speaker Dan Ring
Title ICCV 2009 Review.
Slides Dan's Slides
10th December 2009
Speaker Prof. Anil Kokaram, Dr. Naomi Harte
Title Scientific Writing Forum.
Slides Naomi's Slides,
Anil's Slides
4th December 2009
Speaker Stephen Adams
Title Binaural synthesis for Virtual Auditory Environments
Abstract I'll be discussing the creation of VAEs using binaural techniques and the practical issues and problems encountered with headphone reproduction of binaural along with possible solutions to these problems.
26th November 2009
Speaker Mohamed Ahmed
Title ICIP 2009 Review
Slides Mohamed's Slides
5th November 2009
Speaker Dr. Edward Jones, NUIG
Title Aspects of DSP Research at NUI Galway
Abstract In this presentation, Dr. Edward Jones will discuss some DSP-related projects currently underway in Electrical & Electronic Engineering at NUI Galway. The talk will cover speech-related projects in robust speech recognition, and audio quality assessment, as well as current biomedical signal processing research including the use of ultra wide band radar for the early detection of breast cancer.
29th October 2009
Speaker Darren Kavanagh
Title Speech Segmentation with Applications in e-Learning
Abstract His talk will present an approach for determining the temporal word boundaries in speech utterances.This approach uses methods such as Dynamic Time Warping, DTW and Principal Component Analysis, PCA.Showcase demonstrations of the Recitell application will be given towards the end of the presentation.
22nd October 2009
Speaker Kangyu Pan
Title Spot Analysis in Microscopy.
Abstract The formations of various memories require the present of different specify mRNP in the neuron cells. The development of fluorescent proteins and high resolution fluorescence imaging allow the biologists to locate the mRNP in a living specimen by the co-localization of the differently labeled protein markers. However, the quantitative interpretation of the labeled proteins is still heavily reliant on manual evaluation. In this talk, a novel shape modeling algorithm will be introduced for automating the detection and analysis of the proteins. The algorithm exploits a Gaussian mixture model to characterize the geometric information of the protein particles, and applies Split-and-Merge Expectation-Maximization alg0rithm for optimizing the parameters of the model.
15th October 2009
Speaker Mohamed Ahmed
Title Bayesian Inference for Transparent Blotch Removal.
Abstract Current blotch removal algorithms model the corruption as a binary mixture between the original, clean images and an opaque (dirt) field. This typically causes incomplete blotch removal that manifests as blotch haloes in reconstruction. The talk will start by introducing a new algorithm for removing blotches. The novelity of this algorithm is in treating blotches as semi-transparent objects. The second part of the the talk will discuss a quantitative approach for accessing the restoration quality of blotch removal algorithms. The idea here is to create a near ground-truth blotch mattes by exploring the transparency property of the supplied Infrared scans of blotches. Finally, the talk will conculude by providing a brief overview on current techniques for separating mixtures of natural images.
Upcoming Speakers
| Date | Speaker(s) |
|---|
| 2nd June | Finian/Ken |
| 16th June | Luca/Felix |
