Research

Current research interests:

Extracting and utilizing clinical decision-supportive knowledge from genetic variant literature
Pathogenicity prediction models for genetic variants
In-depth understanding of BERT and building a domain-specific BERT

Past research interests:

Pedestrian detection

A near-frontal head shoulder detector has been developed to detect pedestrian in crowded environments such as shopping malls and downtown streets. The detector consists of two-stage classifiers. First one is a variant of Viola-Jones style ones, i.e., cascade of rejectors, each replying on LDA projection and rapid-computable local features for the real-time operation. The local features include HOG (histogram of oriented gradients) and HSV color values in rectangular local blocks. SVM is then adopted as the second-stage classifier that aims to reduce the false positives, i.e., non head-shoulder patterns that passed through the cascade due to the limitation of the local features. At the second stage, global cues such as color distribution and shape information that could not be dealt with in the first stage have been exploited as features with figure-ground segmentation. Detection results of the detector are shown in the following video clip (the quality is set to low). Most of the processing time is spent in the multiplication operations in the SVM classifier.

head-shoulder detection_svm_0.mp4

Pattern classification (hyperspectral image classification, credit card fraud detection)
Evolutionary computation and its applications to optimization
Neural networks

Google Sites

Report abuse