Tsung-Han Chan's Research Interests
. My first research interest is to develop blind source separation (BSS) algorithms, based on signal processing and convex optimization, for solving the "mixed pixel problem" in image-related applications. BSS is to extract source signals or signals of interest from measurements without information of how the sources are mixed together. The following are some the applications that I am interested in:
Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) provides a powerful noninvasive tool for evaluating tumor (cancer) vasculature based on contrast accumulation and washout. Such vasculature information that reflects tumor angiogenic activity has high potential utility in evaluating the efficacy of cancer treatment on the basis of functional changes observed. Many malignant tumors show heterogeneous areas of permeability; i.e., the signal at each pixel represents a linear mixture of more than one vasculature permeability with various perfusion rates. Our goal is to separate multiple spatial distributions associated with different efflux rates.
- Application in dynamic fluorescent image analysis :
Fluorescence molecular imaging has great potential for advancing basic research and drug discovery and development, but widespread adoption of optical imaging modality is being held back because of obstacles to truly quantitative imaging of deeper organs, tissues, and targets. Mainly due to the malign effects of light scattering and absorption, each dynamic fluorescent image is delineated as a linear mixture of the anatomical maps associated with different major organs. Our goal is to identify the internal organ or structure from which a molecular signal may have originated, and provide the biodistribution dynamics of the major organs, which can facilitate versatility for studies of orthotropic disease, diagnostics and therapies.
Hyperspectral remote sensing exploits the electromagnetic (EM) scattering patterns of distinct materials atspecific wavelengths, and measures the scattered portion of the EM spectrum from the visible region through the near-infrared overhundreds of narrow contiguous bands. However, due to low spatialresolution of the hyperspectral sensor used, each pixel of the observed spectra comprises multiple spectra fromdistinct materials (endmember signatures). Our goal is to unmix the data into endmember signatures and their corresponding mixing proportions (abundances).
. My second research interest is to develop principled computational algorithms for highly effective, scalable, and robust face recognition systems. Here are some methods for improving recognition accuracy:
Our face recognition system requires a set of face images captured under different ‘single-light- source’ illuminations as the training data. The quality of these training images will have a direct impact on the final face recognition rate. As compared to conventional image acquisition where each face image is captured under a single-light-source illumination, a multiplexed illumination scheme which displays distinct illuminations onto the target simultaneously has been proved highly valuable in noise reduction. However, existing optimal multiplexing schemes are primarily designed for ‘time’ multiplexing, and can only apply to a system requiring a number of acquired images equal to the number of illumination sources. Our goal is to design a time and color multiplexed illumination scheme, where not only the order of the illuminations, but also the colors and number of illuminations, are specifically designed so as to minimize the noise power of the acquired images.
For more information, please check my publications.