Research
All my research experiences are categorized into 3 branches: academic research, practical project and group works.
Researches
My primarly selected acadmic researches are presented in this section.
Single Class Object Classification Based On SURF-SIFT and Locality-Sensitive Hashing Algorithm
In this paper, we accomplished a practical movie poster matching system with our proposed SURF-SIFT local descriptor. The performances of SIFT, SURF and our SURF-SIFT algorithms are tested under different image transformations respectively in high dimensional data query. The result turns out that SIFT is better in descriptors while SURF is merit in keypoints detection. Our SURF-SIFT inherits both advantages and achieves a spectacular advantage in rotation in high dimension data matching. Also, Locality-Sensitive Hashing technique is applied as an approximation of traditional nearest neighbor search method, which significantly improves the system efficiency.
Utilization of DICOM Multi-frame Objects for Integrating Kinetic and Kinematic Data with Raw Videos in Movement Analysis of Wheel-chair Users to Minimize Shoulder Pain
Wheelchair users are at an increased risk of developing shoulder pain. The key to formulating correct wheelchair operating practices is to analyze the movement patterns of a sample set of subjects. Data collected for movement analysis includes videos and force/ motion readings. Our goal is to combine the kinetic/ kinematic data with the trial video by overlaying force vector graphics on the raw video. Furthermore, conversion of the video to a DICOM multiframe object annotated with the force vector could provide a standardized way of encoding and analyzing data across multiple studies and provide a useful tool for data mining.
Combining Convolutional Neural Network with Stacked Convolutional Restricted Boltzmann Machine For Object Recognition
We present a novel deep network structure that combines the Convolutional Neural Network (CNN) and novel Convolutional Restricted Boltzmann Machine (cRBM). This combination is analogous to the conventional Deep Belief Network which takes advantage of the structural connections between stacked RBMs and Multi-layer Perceptrons (MLP). We prove that a stacked cRBM can be expressed mathematically in an identical structure as a CNN, with the essential component of a probabilistic max-pooling process that enforces sparsity and helps learn plausible features. We demonstrate superior performance on two standard datasets over current state-ofthe-art methods.
Measuring the Face - Face Recognition of Hereditary Disorders
Hereditary disorders often result in abnormal placement of particular facial features. Because of this, certain genetic diseases can be diagnosed by identifying abnormal feature placement. We developed a robust algorithm for facial feature extraction consisting of three main stages: face detection, extraction of facial landmarks, and evaluation of differences between ethnicities. Two kinds of algorithms, Gaussian-like and Harr-like with AdaBoost, are used to localize face position. The Gaussian-like method failed to acquire face position accurately but did well at extracting the edges of facial features. The Harr-like with AdaBoost algorithm does quite well in detecting faces, and is insensitive to illumination, facial accessories and expression.
A template Matching Approach for Segmenting Microscopy Images
The system aims to segment microscopy images based on a supervised statistical PCA (Principle Component Analysis) modeling method. The general idea is to train a statistical template in terms of the shape and texture of target object structures by utilizing input sample images. This template will be used in the segmentation of a given test image by maximizing the normalized cross-correlation between the template and the test image. The ultimate goal of this system is to achieve the broad application in segmentations in multiple types of microscopy images. The entire system can be divided into two main parts: the training part and the test stage.
MCM
Mathematical Contest in Modeling (MCM) held by COMAP, Inc., USA, challenges teams of students to clarify, analyze, and propose solutions to open-ended problems. The contest attracts diverse students and faculty advisors from over 500 institutions around the world.
A Novel Design of Snowboarding Halfpipe
How to maximize vertical distance and body twist in the air by designing halfpipe shape and modify dominant parameters while taking into account of practical requirements is the key problem in modeling. An effective plan is to first analyze snowboarder’s trajectory, and then find out how factors contribute to the performance of snowboarder.
Air resistance and sliding friction are assumed to be the two major kinds of resistances in modeling. We begin by designing a halfpipe of which the transition is cyclonical shape based on the combination of physical analysis and mathematical theory of Brachistochrone Curve. Then we build an initial Maximum Twist Model based on the theory of goal function optimization, which can also well reflect the transfer of bioenergy to kinetic energy and rotational energy. A rigid-body model and work out the concrete coefficients of snowboard course structure. The optimal parameter includes takeoff angle being 82° and edge height being4.3m, and the maximum total rotation degree in the air is1260°.
Finally, we developed a detailed test system to test our models and to quantify the differences between different halfpipe shapes. Shape parameter limits and recommended standards for snowboard course are provided. We also use existing halfpipe standard to validate the reliability, accuracy and stability of our model.
Projects
In this section, my works are mainly focusing on solving practical problems by programming. At some sense, they are highly related to software engineering.
Image Post Processing System of CT
My internship at Philips and Neusoft Medical System Co. Ltd. was a breakthrough in my undergraduate study, leading me deeper into the field of medical image processing. At my sophomore year, I worked on the project named “Image Post-Processing System of CT” (IPSC), the aim of which was to develop visualization software of commercial CT to view commercial CT images and do serial operations of CT images, such as sharpness, blurring, removing noise, edge extraction, quantitative measurements and so forth. I concentrated more on reading raw CT DICOM data and then converting them into JPEG, designing the user-interface of arranging images, complimenting basic image processing methods, and so forth using C#. This placement not only increased my knowledge in image processing but also made me to have a practical understanding in the field of biomedical applications particularly in medical image systems.
Source Codes Download (Missing!)
Birth Defects Prenatal Referral Information System
Along with the progress made in diagnostic technology, the possibility of prenatal diagnosis for common defects has greatly increased. In fact, prenatal diagnosis has become the last resort for reducing birth defects. However, currently in Liaoning province, prenatal diagnosis has not been generalized. Overall, the current level of prenatal diagnosis is still low in Liaoning Province, an urgent need for technical propagation, training and dynamic monitoring. In terms of content, the need for prenatal consultation of difficult cases includes video images and cases with positive family history. Among these, genetic diagnosis features complex diagnosis, long period, high risk, multiple procedures, in addition with the poverty caused by the frequent medical diagnosis and treatment. These have necessitated the provincial government, hospitals at different levels, laboratories and patients to use an information networks for communication. To establish such an information platform is expected to propagate diagnostic techniques in time and to achieve the standard of prenatal diagnosis reference.
VeeWool Job Searching and Freelance Website Development
The project scope entails designing and developing a website to help Chinese student with abroad education experience finding jobs at China along with freelance jobs availabilities. This website is a combination of job searching and freelance website such as Odesk or Freelance. Thus, we have to provide both the functions in job searching and hire, and also the ones in freelance transaction controls. Besides, the jobs include full-time, freelance, internship, part-time and other approachable job types. Translation, abroad school application consultation and editing are the typical options for freelance jobs.
Applied Techniques:
• CSS+DIV layout (driven from PSD template)
• HTML • Javascript • J2EE • MySQL • Ajax • JSTL • MVC • JDBC
• Demo is available on www.veewool.com.
Function Specifications Download
Personal Information Management System
This project was conducted during a summer internship at Philips and Neusoft Medical System Co. Ltd when I was in college. It was developed in C++ using Visual Studio to enable users managing their daily schedules such as meeting, birthday, holidays and setting alarm clock.
Group Works
The followings are selected group works during my studies in different periods.
Brain-Computer Interface Technology to Address Human Paralysis
In this paper, we review brain-computer interface (BCI) technology as applied to human paralysis. We give detailed descriptions of human brain function, electrode implantation to gather signals, and neuro signal processing. The challenging issues remaining in this field and future developments in BCI technology are also discussed at the end of this paper.
Design and Simulation of Somitogenesis Mechanism in Zebrafish Model
Vertebrate animals will experience an obvious characteristic stage in their developing embryo, that is Somitogenesis. This process happens after the initialization of gastrulation and somitogenesis will form a repetitive structure in an anterior-to-posterior direction during body axis elongation.
In the research of somitogenesis,we choose to study the previous models and theories applied to simulate the mechanism of zebrafish, which is a typical species for many researchers. We mainly referred to four different models and analyze the advantages and disadvantages of them, which are respectively “cell oscillator model”, ”reaction-diffusion typemodel ”, ”cell-cycle model”and“clock and induction model”.
The Exploration of Type 2 Diabetes Mellitus and Related Proteins Based on P-Value
Diabetes mellitus, often simply referred to as diabetes—is a group of metabolic diseases in which people have high blood sugar, either for the body does not produce enough insulin, which is called Type 1 diabetes, or because cells do not respond to the insulin that is produced, which is called Type 2 diabetes. In this paper, we will investigate Type 2 diabetes and 13 kinds of proteins, which could be have some relationship with Type 2 diabetes mellitus through the given experiment’s data by Tu/e and through the mathematical method which is to calculate the P-Values of the data and then use the results to propose our assumptions. Finally, we will do some plumb researches through the Internet and some bibliographies to demonstrate our previous hypothesizes.
Introduction to Computer Vision
The visual system is our most significant sense, which is a signal processing sytstem with spectacular performance. It is capable of extracting structural information from the visual scene, with both astonishing capabilities and real-time speed. The task of the system is of great importance to help us survive in the world, therefore, it is necessary to start a detailed anaylsis of 2D image as projection of the 3D world. We are going to give a summary of the properties of three main parts(retina, Lateral Geniculate Nucleus and cortex) in human visual system. At the last part of this summary, we will have an insight into cortical columns, which acts as one main structure of cortex.
Fingerprint and Blood Vessel Curvatures Analysis
Fingerprint recognition has been increasingly used to recognize personal identification, such as ID card, fingerprints hard disk and so on. Great improvement has been achieved in the fingerprint sensing technology and automatic fingerprint recognition algorithms.At the same time, the vessel segmentation maintains the importance in pattern recognition. However, the fast,reliable and computerized matching of fingerprint images is a remarkable problem in pattern recognition has not yet received a complete solution.This paper focus on solving the basic problems in fingerprint and vessels segmentation of fundus.
Image Debluring
In most cases of our life,we are always confronted with such problems that the images are degraded by many factors. When we use a camera,we want the recorded image to be a faithful representation of the scene that we see--but every image is more or less blurred.Thus,image deblurring is fundamental in making pictures sharp and useful. In this paper we apply a novel approach, Space-Scale Algorithm, to produce a high quality image. We adopted an image with license number of a vehicle as the object. The image with license number is blurred and we will also deblur this image by by some other image processing methods. The main algorithms we considered in this paper are Space-Scale Algorithm, Point Spread Function Deblurring Algorithm, Wiener Filter Deblurring Algorithm, Poisson Algorithm and Richardson-Lucy Algorithm. After processing and comparison, we obtained the satisfied results and our conclusions.
Testimonials
"My heart is in the work."
Andrew Carnegie