Scholarships 2011 Scholarship from Queen Mary to study in London.
2014 Magdalen Award from China Oxford Scholarship Funds.
Computer skills Programming Language: Matlab,Python, C++ (OpenCV, OpenGL), JAVA

谢伟迪

Weidi XIE

  • Tel: +44 07598730577
  • Email: weidi.xie@wolfson.ox.ac.uk

不怕世界上有如此多的未知,但我害怕自己一无所知。
Education P.h.D Candidate,University of Oxford Topic: Machine Learning and Biomedical Image Analysis
Supervisor: Prof. Alison Noble
Master of Science, University College London Major: Computer Graphics,Vision & Imaging (Computer Science) Bachelor of Science, University of London, Queen Mary Major: Electrical Engineering (Telecommunication Engineering)
First Class Honours
Bachelor of Science, Beijing University of Posts and Telecommunications
Major: Electrical Engineering (Telecommunication Engineering)
GPA : 87/100
Bachelor Thesis Seamless Image Cloning Supervisor: Prof. Ebroul Izquierdo
Description: The project aimed at implementing a seamless cloning image editing tool by solving Poisson equation. Users can import(clone) transparent and opaque source image regions into destination image or modify the appearance of the image within a selected region seamlessly.
Keywords: Image Processing, Poisson Cloning
Masters Thesis Handwriting Authorship Recognition Supervisors: Dr. Lewis GRIFFIN & Dr. Andrew NEWELL
Description: In this thesis, we used oriented Basic Image Features (oBIFs) to extract features, and described how Non-Negative Matrix Factorization can be used to solve the inverse problem. Our method aimed at extracting ’Handwriting- Style Subspace’ from training samples, then, identified the authorship of input words with machine learning approaches. This new method can be useful in forensic science.
Keywords: Computational Vision, Image Processing, Data Analysis, Machine Learning Non-negative Matrix Factorization
Project Experience Cell Tracking In order to visualise the transcriptional bursts in single Dictyostelium cells, I developped a cell tracking system with GUI for MRC Lab for Molecular Cell Biology at UCL.
Region-based level-set method were used, each cell in the first frame was manually assigned to one level-set function by clicking cell center.
Put strong penalty on overlap level-sets to keep them seperate.
Put penalty on cell size changes.
Then recursively initialized the next frame with segmentation result from previous frame.
Finally, created seeds and use watershed to make an improvement.
Robust Text Detection and Recognition from Natural Images
localize text candidates with two methods, which are Stroke Width Trans- form (SWT) and Maximally stable extremal regions (MSER).
In order to classify character and non-character, build unsupervised deep network with denoising autoencoders, then the pre-training of such archi- tecture was done one layer at a time.
Fine-tuned the network with small number of labelled data.
After detection of text, recognition was based on a deep Convolutional neural network, which is robust to noise, scale variance and translation.
The algorithm has been tested on ICDAR 2013 datasets, and has shown state-of-art result.
Will submit to the 2015 International Conference on Document Analysis and Recognition (ICDAR) competition.