My research interests broadly involve Computer Vision, Machine Learning, and Information Retrieval. I have been working on the problems involving both spatial (image) and temporal (video) domains:

  1. Automatic face analysis
  2. Unsupervised commonality analysis
  3. Image & information retrieval



Unsupervised Temporal Commonality Discovery

This study explores a relatively unexplored commonality discovery problem for temporal domain. We term this problem Temporal Commonality Discovery (TCD) . We propose a Branch and Bound (B&B) algorithm to efficiently find the global optimum with designed bounding functions for L1 , L2 , and Χ2 distances. Our algorithm is general and can be applied to any feature that has been quantified into histograms. To the best of our knowledge, TCD is the first work that addresses unsupervised discovery of common events in videos.

[ more ]



Gender Ideitification from Unaligned
Face Images

Rough face alignment leads to suboptimal performance in face identification systems. This work presents a novel approach to identify genders from facial images without proper face alignments . We capture and match sets of unaligned face images using linear subspaces and set classification.The proposed approach showed competitive results with state-of-the-art methods that consider face alignment.

[ more ]


What pattern are common?


Common Pattern Discovery & Image Co-segmentation

This work presents an efficient method for finding common visual patterns among a set of images. We tackle he Common Pattern Discovery (CPD) problem as dense cluster discovery in the feature space. Unlike many relevant studies, CPD allows multiple common patterns to occur multiple times .


We further apply a novel Gibbs energy model on the CPD results to refine the segmentation bonudaries, which we termed MOMI-cosegmentation .Experiments showed that our approach outperforms state-of-the-art co-segmentation methods.

[ more ]


Differences between eigenvectors and canonical vectors


Set-Based Face Recognition

This work proposes to recognize faces using multiple instead of a single face image. The idea is to capture various face information through an image set. We extended the concept of canonical correlation to set matching problems in the RHKS. In particular, we proposed two face recognition frameworks: Kernel Discriminant Transformation (KDT) and Component-Based Constraint Mutual Subspace Method.

[ more ]


Which region should we consider as query image?


Region-Based Image Retrieval

Relevance feedback and region-based image retrieval are two widely used methods for enhancing content-based image retrieval systems. This work combines these two approaches to better capture human's perception. The proposed system takes user feedbacks to model their particular interests. Then, we measure the similarities between image regions using a proposed Group Biased Discriminant Analysis (GBDA) based on user feedbacks.

[ more ]


License plate identification system flowchart


License Plate Recognition

In this project, we implemented a license plate recognition system under a wireless framework. Clients (or users) can simply capture a picture of license plate with their hand-held device and retrieve the information of the vehicle owner. The proposed system consists of three primary processes: license plate registration, text segmentation and Optical Character Recognition (OCR). The police could benefit from this fully automatic and serve-on-demond license plate identification system.