|
|
|
|
|
|
Vinod Chandran |
||||||||||||||||||||||||||||||||
 
Profile
Qualifications, Career history and Professional and Group Associations QualificationsPhD (Washington State), MSCompSci (Washington State), MSElectEng. (Texas Tech), BTechElectEng (IIT Madras) Career History and biography
Professional and Group Associations
Interests and community serviceTechnical Program Chair, Workshop on Signal Processing and Applications (WoSPA), 2002 
Research
Research areas and external collaborators Research AreasWithin the broad field of Signal Processing and Machine Learning, Associate Professor Vinod Chandran and his research team has defined nine main research areas:
Signal Processing
A speech waveform
A spectrogram
Phase (bi-phase) of the bispectrum integrated along a diagonal line in bi-frequency space (hence a function of only one frequency) as a function of frequency and time
Associate Professor Chandran and his research team have made original contributions to the theory and application of signal processing. In particular, they work on non-linear and high order statistical(HOS)techniques. Theoretical contributions include the derivation of principal domains of computation for HOS of arbitrary order, derivation of statistics of the bispectrum in the presence of leakage and estimation of the statistics of bicoherence and tricoherence. A new method of extracting similarity transformation invariant features was discovered and developed by A/Prof Chandran in 1993. The image on the left shows it applied to speech data, demonstrating potential for use of bispectral phase information which shows structure unlike Fourier phase which is random. HOS features are more robust to additive noise than spectral or cepstral features. HOS techniques have been applied by A/Prof. Chandran and his team to Speech, Ground Penetrating Radar, EEG, and Heart rate variability (HRV).
Image Processing
4 facial images obtained at poor resolution typical of frames from surveillance video. They are appropriately combined using optical flow and super-resolution techniques developed by PhD student Frank to obtain the image on the right which has better resolution. Improvement in resolution and quality leads to improvement in face recognition – human and machine Associate Professor Chandran and his team have investigated image processing techniques to enhance facial images from surveillance video using super-resolution methods (image on the left), extracting features from facial images for automated face recognition, fractal iterated function systems, Gabor filter based multi-resolution decompositions, part-face feature modelling, eye tracking to detect regions of interest (ROI) and ROI coding in JPEG2000, and higher order spectral invariant feature extraction from Radon transform projections. They have applied new image processing techniques to face recognition, clinically significant maculopathy detection in fundus images, logo recognition in document images, circuit diagram interpretation in document images and virus classification from electron microscope images.
Pattern Recognition
Pseudo-colour images of regions from a larger sonar image after several processing steps Associate Professor Chandran and his research team have developed a number of innovative techniques in extracting robust features from higher order spectral line integrals and entropy parameters, and their use with Gaussian mixture models and Support Vector Machines. They have applied the new methodologies to biometrics [speaker verification, face recognition], biomedical signal and image processing [onset of epileptic seizure detection from EEG, heart condition monitoring using HRV, maculopathy detection in fundus images], radar signal processing [detection of coal rock interface using Ground Penetrating Radar] and sonar image processing [detection of sea mines – image on the left]. Most recently, A/Prof. Chandran and his team are working new ideas on multiple classifier systems with feedback and iterative chaotic transformation based binary feature extraction. They are intended for use with multimodal biometric systems and cryptographic key generation. The image above shows pseudo-colour images of regions from a larger sonar image after several processing steps: adaptive Wiener filtering, approximately matched filtering and adaptive threshold operations to clip extreme-valued pixels. The top left image is of an underwater mine. The top right image contains only noise and the bottom image is that of an impostor or other underwater object. These images are now suitable for feature extraction and machine learning algorithms that can be trained to detect underwater mines. Biometrics
Three dimensional surface representation of a human face
Flow chart showing how 2D and 3D features are combined to improve face recognition Associate Professor Chandran and his research team have particular expertise in the areas of face recognition and speaker verification.
Three dimensional (3D) face data obtained using laser range finders or stereo cameras is of current research interest because of its robustness to lighting variations and facilitation of pose determination for alignment. Two PhD theses investigated methods of using 3D faces for recognition and combining them with 2D faces for improving performance. A third PhD investigated the use of super-resolution technique to enhance faces in surveillance video frames such that they can be used for recognition. In the area of speaker verification, Associate Professor Chandran and his team have compared the performance of higher order spectral (HOS) invariant features with the more traditional Mel-frequency Cepstral Coefficient (MFCC) features in conjunction with GMM classifiers. The HOS features retain some Fourier phase information and are more robust to noise. Current work in biometrics includes PhD theses on prompted text-dependent multiple classifier speaker verification system and a biometrics based cryptography system. Face Verification
Gabor filters are used to decompose a facial image into different frequency bands and spatial regions and the effectiveness of each in machine recognition of faces is analysed in order to obtain the best weighted combinations. Associate Professor Chandran and his research team have investigated the performance of face verification using Gabor filters and multi-scale decompositions to determine the proportion to which different parts of the face and different frequency bands or scales contribute to their accuracy. For 3D faces, the nose tip region, and for 2D faces the bridge between the eyes, are good discriminating areas quite unaffected by expression variations and experimental results by Dr. Jamie Cook highlight this.
Other members of the team have investigated part face based feature modelling and fusion of 2D and 3D features to improve performance, fractal features for face recognition, person tracking and super-resolution for face recognition from surveillance video and effect of super-resolution on face recognition using Eigenfaces and Elastic Bunch Graph Matching. Signature Verification
A pen-tablet system being used for on-line signature verification. Associate Professor Chandran developed a technique using time-varying higher order spectral features and Gaussian Mixture Models for online signature verification. He has also authored HOS and Classifier libraries in C and participated in the first signature verification contest held as part of the International Conference on Pattern Recognition (ICPR), Hong Kong, in 2004. The system is fast and does real-time classification.
An online demonstration prototype was implemented by Associate Professor Chandran and research team at QUT and some experiments were performed on in-house data collected from family members by A/Prof. Chandran. Higher Order Spectral Analysis
Plot showing a radial line through bi-frequency space and the triangular non-redundant region of computation of the bispectrum. Phases of bispectral integrals along such lines satisfy similarity transformation invariance properties useful in pattern recognition Associate Professor Chandran has made seminal contributions to the theory and practice of higher order spectral analysis stretching back to 1988. He has published original work on 2D bispectral analysis, statistics of the bispectrum in the presence of leakage, a general procedure to derive principal domains of higher order spectra and statistics of the bicoherence and tricoherence, a new similarity transformation invariant feature extraction technique using bispectral integrals and an extension of the feature extraction technique to 2D images, all in Transactions of the IEEE.
A/Prof. Chandran and his research team at QUT have developed and applied higher order spectral techniques in many areas such as speaker verification, radar signal processing, sonar signal processing, document image processing, biomedical signal and image processing and biometric based cryptographic key generation. Shown on the left are plots of the bispectrum magnitude and bicoherence of heart rate variability in a patient with Atrial Fibrillation obtained by PhD student, Chua Kuang Chua. Biomedical Signal and Image Processing
A fundus with damage from diabetic maculopathy. Image processing techniques were used to detect such degeneration that is clinically significant. Associate Professor Chandran’s feature extraction technique has been applied by Chua Kuang Chua and project students under his supervision in the development of a system for detection of clinically significant diabetic maculopathy in fundus images (image on the left).
A/Prof. Chandran and Chua Kuang Chua have applied higher order spectral and nonlinear techniques to analyse and classify biomedical signals such as heart rate variability (HRV) and Electroencephalograph (EEG). They have successfully classified HRV signals into several cardiac condition classes. They have also classified EEG segments into normal, pre-ictal and epileptic seizure classes in a system to detect the onset of an epileptic seizure. A/Prof. Chandran is currently investigating bicoherence and power spectra of EEG signals (image on the left) in the alpha and beta frequency bands during ERD(event related desynchronization) that occurs due to voluntary and involuntary activity. The target applications are rehabilitation of stroke survivors with arm deficit and use in brain computer interfaces. Cryptography
A/Prof. Chandran and PhD student are investigating a method for generating cryptographic keys from biometric information such as that obtained from facial images in video frames. The challenge is in obtaining crisp (free of even a 1 bit error) keys from inherently fuzzy biometric data or features. A unique technique based on chaotic iterative transformation and bispectral integrals has been developed by them Associate Professor Chandran developed an iterative transformation based on bispectral integrals for feature extraction from one-dimensional signals in 2006. The method is an extension of his invariant feature extraction method developed during 1990-93. The central idea is to extract a sequence of features such that for small changes in the input the sequence will diverge while for the same deterministic input sequence the output will always be the same. The features could then be binarized and the bits used for pseudo-random number generation and cryptographic key generation or release.
A/Prof. Chandran and Brenden Chen are currently developing the technique for cryptography and applying it to faces in video and 3D form.
Teaching
Teaching areas and achievements and units taught Teaching areas
Grants
Funding and selected list of awarded projects Funding First Chief investigator on 2 ARC Discovery Grants (2001-4,2004-7) and 2 large international research contracts (1997-99,1999-2001) and co-Chief Investigator on an ARC Linkage (2007-10) and Industry and ARC Small Grants totalling more than $1 Million. Techniques to use stereo vision for improving person identification systems based on face recognition
A three dimensional surface representation of a human face where height is indicated by colour. Role: First Chief Investigator
Speech recognition using source dependent information from phases of higher-order spectra of quasi-periodic
Gaussian mixture models obtained for higher order spectral features Role: First Chief Investigator
The image shows Gaussian mixture models obtained for higher order spectral features (four plots on the left) and Mel frequency cepstral coefficient (MFCC) features were compared by A/Prof. Chandran and Dr. Daryl Ning. The study indicated that HOS features can yield complementary information to improve speaker verification and are also more robust to additive noise Speech recognition using source dependent information from phases of higher-order spectra of quasi-periodic speech components
Plot showing the magnitude of the bispectrum for a speech segment Role: First Chief Investigator
The plot shows peaks at bi-frequency pairs that involve harmonics of the pitch frequency. This is typical of voiced speech segments. For such quasi-periodic components the phase might carry potential source (or speaker dependent) information. The bi-phase, being translation invariant, is suitable to extract such information. Multimodal speech processing and its application to speaker recognition
Lip movements can be analysed using video/image processing techniques and used to recognize speech in addition to the audio signal. The multi-modal approach can yield better accuracy especially in high audio noise environments.
Role: Chief Investigator
Chief Investigator: Sridharan, Chandran Participating agencies: ARC Amount: $11,000 Period: 2000 3D Face Identification and Verification
A three-dimensional face is decomposed using Gabor wavelets into spatial and spatial frequency regions. Classifiers trained on each region can be combined in a weighted manner to achieve recognition rates. Role: Investigator
Detection, Classification and Identification of Objects in Cluttered Acoustic Images
Pseudo-colour images of regions from a larger sonar image after several processing steps Role: Chief Investigator at QUT on the Research Contract
The images shown here are pseudo-colour images of regions from a larger sonar image after several processing steps: adaptive Wiener filtering, approximately matched filtering and adaptive threshold operations to clip extreme-valued pixels. The top left image is of an underwater mine. The top right image contains only noise and the bottom image is that of an impostor or other underwater object. These images are now suitable for feature extraction and machine learning algorithms that can be trained to detect underwater mines.
Speech and Image Recognition Technologies
Lip movements can be analysed using video/image processing techniques and used to recognize speech in addition to the audio signal. The multi-modal approach can yield better accuracy especially in high audio noise environments Role: Co Investigator
Image Recognition Technologies
Results of a skin detection algorithm developed at QUT Image Recognition Technologies
Role: Co Investigator
These images show the results of a skin detection algorithm developed at QUT. Skin regions are shown with grid-lines. Skin detection is used to locate faces and is a step towards automated face recognition/verification and person tracking.
Importance Prioritized JPEG2000 Image Coder
Results obtained by PhD student Anthony Nguyen with regions marked by ellipses Role: Co Investigator
Not all regions in an image are equally important to a viewer. The regions of interest (ROI) can be found using eye movement data and/or using machine learning algorithms. These images show results obtained by PhD student Anthony Nguyen with regions marked by ellipses. The ROI information can be used in an image compression algorithm such as JPEG2000 such that these regions are encoded with better fidelity than other regions.
2D Face Recognition Trials
Face recognition systems typically use such rectangular regions to extract features because they avoid regions that vary a lot for images of the same individual – such as hair 2D Face Recognition Trials
Role: Co Investigator An automated face recognition system was implemented at QUT. Shown above is a result after the face location step. Face recognition systems typically use such rectangular regions to extract features because they avoid regions that vary a lot for images of the same individual – such as hair.
Supervision
Selected list of student projects Within the broad field of Signal Processing and Machine Learning, Associate Professor Vinod Chandran’s students have undertaken following funded research projects:
Principal Supervisor of research theses:
A Decompositional Investigation of 2D and 3D Face Recognition
A three-dimensional surface representation of a human face with height indicated by colour Name: James Cook
A three-dimensional surface representation of a human face with height indicated by colour. Gabor filters were used to decompose a facial image into different frequency bands and spatial regions and the effectiveness of each in machine recognition of faces was analysed in order to obtain the best weighted combinations. Hybrid 2D and 3D Face Verfication
A flow chart showing how 2D and 3D features are combined to improve face recognition Name: Chris McCool
Part-face methods and feature modelling were investigated in this thesis. Robust Thin Layer Coal Thickness Estimation using GPR
A Ground Penetrating Radar pulse signal is transmitted by one antenna and the beam reflected by various layers is received by another Name: Andrew Strange
A Ground Penetrating Radar pulse signal is transmitted by one antenna and the beam reflected by various layers is received by another. The received signal cannot be detected by standard matched filters with an accurate estimate of the delay which is proportional to the layer thickness when the layer thickness is small as would be the case when a coal seam has been stripped nearly down to the rock. A pattern recognition approach with higher order spectral features was adopted in this thesis to train a system that will detect when the interface is within a specified range. It can train to the different pulse shapes that can arise at different thicknesses and adapt to variations in system parameters such as the dielectric constant. Fractal Techniques for Face Recognition
Fractal parameters were investigated as features for face recognition Name: Hossein Komleh
Fractals satisfy self similarity at various scales and arise in the study of non-linear dynamics and chaos. One part of an image may also be self-similar to another and can be obtained by a transformation. A partitioned iterated function system (PIFS) attempts to encode an image using parameters of such transformations and forms the basis of fractal encoding of images. In this thesis, fractal parameters were investigated as features for face recognition. Classification of Viruses using HOS
Virus particles in electron microscope images often appear as textured blobs and it is hard to tell one virus from another by visual examination of a single particle or even a group Name: Hannah Ong
Virus particles in electron microscope images often appear as textured blobs and it is hard to tell one virus from another by visual examination of a single particle or even a group. However, a large number of such particle projections are obtained from even a single specimen. Such projection ensembles have been used in electron microscopy to obtain three-dimensional reconstructions by iterative alignment and back-projection methods. In this thesis, a method for automated classification of viruses from such ensembles is investigated and developed. This technique does not require reconstruction and can work with smaller ensembles while testing. The method is based on similarity transformation invariant features obtained from bispectral integrals of Radon transform projections. ROI coding in JPEG2000 for Surveillance Applications
Results obtained by PhD student Anthony Nguyen with regions marked by ellipses Name: Anthony Nguyen
Not all regions in any image are equally important to a viewer. The regions of interest (ROI) can be found using eye movement data and/or using machine learning algorithms. The images shown are results obtained by PhD student Anthony Nguyen with regions marked by ellipses. ROI coding was implemented within the JPEG2000 framework and applied to large format aerial surveillance imagery for defence applications in this thesis. Video Event Detection
Frames from a video recording of a person walking in the laboratory and their binarized versions showing the person segmented from the background Name: Chaitanya Gurrappu
The image shows frames from a video recording of a person walking in the laboratory and their binarized versions showing the person segmented from the background. Such segmentation can be error-prone because of shadows and varying illumination. It is a pre-processing step in the analysis of video to detect events such as a person writing graffiti on a wall or ticket vending machine. This thesis investigated the use of hidden Markov models with features extracted from the person segments for simple video event detection. Document Anlaysis and Recognition using Image Processing Techniques
Automated recognition of a logo can help sort the documents for proper channelling and archiving Name: Scott Lowther
Official documents and invoices often have logos of organisations such as those featured in the image. Automated recognition of the logo can help sort the documents for proper channelling and archiving. This thesis investigated methods of segmenting and skew-correcting logo regions and applied bispectral invariant features to the task of automated logo recognition in documents. The particular approach is suited because the logo regions can appear with small scale changes or shifts and can be corrupted by noise. Associate Supervisor of research theses:For detailed information on these theses, please refer to the profile of the principle supervisor in each case. Very Low Bit-rate Multimedia Coding
These images were taken in the laboratory and their binarized versions that show a background in black and the foreground in white Name: Darren Butler
These images were taken in the laboratory and their binarized versions that show a background in black and the foreground in white. The foreground moves (hence the Frisbee image) whereas the background is static. However, the segmentation task is complicated because lighting can change and the background changes without any motion. PhD student Darren Butler developed an algorithm for fast background segmentation that is adaptive and robust to such lighting changes. Model Based Trainable Speech Synthesis and Applications
3D spectrum of synthesized speech Name: John Dines High quality Audio Compression using the Wavelet Transform and Optimal Bit Allocation
The power spectrum of an audio signal and the error (residual) spectra from the method proposed in this thesis and the LPC method of audio compression Name: Daryl Ning Principal supervisor: A/Prof M. Deriche
The image shows the power spectrum of an audio signal and the error (residual) spectra from the method proposed in this thesis and the LPC method of audio compression
Face and Lip Tracking
Techniques to track the human face and lips in video for use in multi-modal speech recognition and speaker verification Name: Simon Lucey Principal Supervisor: Prof. S. Sridharan
This thesis investigated techniques to track the human face and lips in video for use in multi-modal speech recognition and speaker verification Robust Speaker Recognition
Investigating techniques in the feature domain Name: Jason Pelecanos
This thesis investigated techniques in the feature domain and proposed a feature warping method to improve speaker verification based on Gaussian mixture models Robust Audio Coding
Investigating methods of audio coding that are suited to covert surveillance Name: Michael Mason Principal Supervisor: Prof. S. Sridharan
This thesis investigated methods of audio coding that are suited to covert surveillance as conducted by Police for criminal investigation Person Authentication using Audio and Video Information
Investigating the use of lip features Name: Tim Wark
This thesis investigated the use of lip features (area based and contour based) and hidden Markov models in person authentication
Research Projects Currently being Supervised
Cryptographic Key Generation from Biometrics using an Iterative Chaotic Bispectral Transform
Histograms of bit errors in keys generated for the true identity (top) and impostors (bottom) Name: Brenden Chen
The plots on the images show histograms of bit errors in keys generated for the true identity (top) and impostors (bottom). It is apparent that the top histogram peaks at 0 bit error and is quite sharp while the bottom histogram is distributed over the entire key length and has a mode close to half the key length. Keys were generated from 3D faces using a procedure involving bispectral integrals and iterative transformations. The technique is being investigated in this thesis and will be applied to facial images in video. It is attempting to improve the security of electronic transactions and eliminate weak links such as reliance on passwords to protect locally stored private keys. Application of Higher Order Spectral and Non-linear Techniques to Biomedical Signals
the distribution of the magnitude of the bispectrum (left) and the bicoherence (right) obtained from a segment of a heart rate variability (HRV) signal in bi-frequency space Name: Chua Kuang Chua
The images show the distribution of the magnitude of the bispectrum (left) and the bicoherence (right) obtained from a segment of a heart rate variability (HRV) signal in bi-frequency space. Patterns and the extent of randomness (entropy) in this space can be used for feature extraction and classification such that cardiac conditions can be diagnosed or confirmed. Other non-linear methods such as Poincare plots are also useful and this thesis investigates a range of non-linear methods for pattern classification problems in biomedical signals such as the HRV, ECG and EEG. Tracking and Recognition of Persons in Video Surveillance Imagery
Sequence of video frames with boxes marked around persons Name: Simon Denman
The image is a sequence of video frames with boxes marked around persons. The objective is to track persons in frames of surveillance video as they move about regardless of changes and occlusion by other persons or objects. A number of techniques including the particle filter and scalable condensation filter were investigated and the methods developed were benchmarked using standard video databases. Tracked objects and persons were also used in a framework to identify events such as the leaving of baggage unattended. A Holistic Approach to Texture Analysis and Classification
Sand particles of varying sizes Name: Ronald Elunai
The images are of sand particles of varying sizes. In this thesis, such images are treated as ‘particulate’ textures. A method for estimating particle size distributions is developed based on edge detection, run length distribution and correction factors to account for gaps and differences in shape. It is compared with an established method based on autocorrelation and with ground truth obtained via manual sieving. These distributions are useful in sedimentology for studying transport phenomena on an ocean or river bed. They can also be applied to aggregate mixtures used in the construction industry. Texture analysis and classification in the ‘grey’ area where object or particle properties are still relevant is the subject of this thesis. Super-resolution Techniques with Application to Face Recognition
4 facial images at poor resolution are combined using optial flow and super resolution techniques to obtain better resolution Name: Frank Lin
The image shows on the left 4 facial images obtained at poor resolution typical of frames from surveillance video. They are appropriately combined using optical flow and super-resolution techniques developed by PhD student Frank to obtain the image on the right which has better resolution. Improvement in resolution and quality leads to improvement in face recognition – human and machine. This thesis investigated a new super-resolution technique based on optical flow that is suited for face images and evaluated its improvement of face recognition using Eigenface and Elastic Bunch Graph Matching Methods. Human Activity Recognition
Two wide-angle surveillance video camera frames with persons indicated by red and green marks Name: Andrew Yau
The image shows two wide-angle surveillance video camera frames with persons indicated by red and green marks. The aim of this PhD is to investigate techniques for the recognition of human activity in surveillance video. It will develop semantic analysis techniques starting from segmented and tracked objects. Multimodal Biometric Sequential Decision Systems
Diagram depicting a multi-classifier architecture in a sequential framework with multiple attempts allowed at each classifier. Name: Vishnu Priya Nallagatla
The diagram is depicting a multi-classifier architecture in a sequential framework with multiple attempts allowed at each classifier. False rejections are indicated in orange. A closed form formula for false acceptances and false rejections relating them to the number of classifier stages and the number of allowed attempts was derived by A/Prof. Chandran for the statistically independent case. This thesis will investigate the architecture in the context of text-dependent speaker verification and develop it for the more general case.
Publications
Selected list of publications Selected List of PublicationsFor more publications eprints: http://eprints.qut.edu.au/view/person/Chandran,_Vinod.html
| |||||||||||||||||||||||||||||||||
