Image&Video Understanding,...
| |
| Home |
| Resume |
| Research |
| Publications |
| Albums |
| Contact Me |
My Research (under construction) |
||
| Areas of Interest | ||
My research fields include image & video processing, pattern recognition, machine learning and computer vision. Currently, I am mainly engaged in image&video understanding, e.g., remote image processing, object/event detection/recognition, video surveillance, etc.
During my Ph.D. study, I have proposed many novel and useful algorithms for iris biometrics. These state-of-the-art algorithms involvs in several important aspects of iris biometrics, such as image acquision (human computer interface), object detection, image segmentation, spoof detection, feature extractin/selection as well as classification. Most of them have been successfully applied in working iris recognition systems (e.g. IrisKing, IrisGuard, ...), which greatly advances the performance of iris recognition. Please refer to Biometrics for more details.
NOTE: I have received my Ph.D. degree in 2010, and my research interests have changed to image&video understanding, such as remote sensing image processing, object/event detection/recognition, video surveillance, etc. Therefore, any questions related to iris recogniton would not be fully supported.
Machine Learning - Adaboost
Adaboost was introduced in 1995 by Freund and Schapire, see Fig. 1. It is a well-known large margin learning algorithm that can select a small set of the most discriminative features (from a candidate feature pool), and combines them into an ensemble classifier. Viola-Jones's work (Robust Real-time Object Detection) made Adaboost learning world facous in the community of computer vision and pattern recognition.
Figure 1. A genelized flowchart of Adaboost learning. It involves three key modules, namely the weak learner, the component classifier and the re-weighting function.
To better understand the machanics of Adaboost learning, Figure 1 gives a genelized flowchart of Adaboost learning. We can see that it consists of three key modules, namely the weak learner, the component classifier and the re-weighting function.
a) The Weak Learner: The weak learner is essentially the criterion for choosing the best feature t on the weighted training set.
b) The Component Classifier: The component classifier outputs the confidence of a sample being a positive based on its t value.
c) Sample Re-weighting: Sample re-weighting enables that the subsequent component classifier can concentrate on the hard samples by assigning higher weights to the samples that are wrongly classified by previous component classifiers.
|