My Research
Current Research
My current research at CMRR is focussed on both methods development for diffusion MRI data analysis, and the applications of the methods for disease data analysis. I proposed algorithms for improved reconstruction of white matter fiber parameters from compressed (under-sampled) diffusion MRI data [Neuroimage 2018, MICCAI 2017, MICCAI 2015]. I introduced the concept of Sparse Bayesian Learning to neuroimage data analysis, that can potentially reduce the diffusion MRI scan-time, and at the same time, provide high quality reconstructions.
Sample image showing improved crossing fiber detection, with lower estimation uncertainty
I am also actively working on analyzing the data to identify areas of degeneration in neurological diseases such as Amyotrophic Lateral Sclerosis (ALS), Ataxia, and Traumatic Brain Injury (TBI).
Doctoral Research
Thesis Title: Computational Intelligence Techniques in Visual Pattern Recognition
Advisors: Prof. Prahlad Vadakkepat & Prof. Loh Ai Poh
My Phd thesis was focused on computational intelligence techniques for visual pattern recognition. We proposed novel algorithms for feature extraction, feature selection, and classification using computational intelligence tools. My thesis addressed various issues like varying object sizes, shape variations, complex backgrounds, inter-class similarity and real time performance in pattern recognition, with its application to hand posture recognition. The main contribution of the thesis, namely a novel Bayesian model of Visual Attention and the utilization of visual cortex model for addressing complex background problem in hand posture recognition, was published in the International Journal of Computer Vision (IJCV).
Visual pattern recognition pipeline, showing my PhD research focus
The Fuzzy-Rough Classifier
Fuzzy sets are utilized to handle vagueness. The concept of Rough sets helps to handle indiscernibility. These two theories can be combined to form Fuzzy-Rough sets. Fuzzy-Rough sets, which is a deviation of Rough set theory, is useful for decision making in situations where both vagueness and indiscernibility are present.
We proposed a novel algorithm based on fuzzy-rough sets for the recognition of hand postures and face [IJHR 2010]. The concepts of fuzzy and rough sets are combined to develop a simple and effective classifier. The relevant features in the dataset are selected using a Genetic Algorithm (GA). The Fuzzy-Rough classifier has good real-time performance (more than 3 times faster compared to a standard SVM classifier) and accuracy (equivalent or better than SVM).
We proposed a variation of the above classification algorithm with a novel feature selection algorithm in [ASOC 2011]. The algorithms are applied to different classification problems; cancer classification and image pattern classification. The proposed algorithm provided classification accuracy which is equivalent or better than that provided by SVM, with less computational efforts, and with a good margin of classification.
Hand Posture recognition using neuro-biologically inspired features
Hand posture recognition is done using neuro-biologically inspired features in [5], [6]. The image features are extracted using a computational model of the ventral stream of visual cortex. The real-time implementation of the algorithm is done for the interaction between the human and a virtual character Handy.
Attention based Hand posture recognition against complex backgrounds
The complex background problem in hand posture recognition is addressed by combining the Bayesian model of attention and the computational model of visual cortex.
As the number of available hand posture datasets is limited, we developed two new 10 class datasets namely NUS hand posture dataset-I & II. The datasets can be downloaded and used for academic research purposes free of cost, by citing the IJHR and IJCV papers respectively.
Sample images from NUS Hand Posture Dataset-II