Offline Handwriting Recognition

Offline Handwriting recognition can be used to recognize & index human written documents. As there are many similarities between speech data and handwriting data (both can be treated as sequences), we want to tailor some of the mature techniques in Speech Recognition and tune them to best fitting Offline Handwriting Recognition task. For example, Hidden Markov Model & Recurrent Neural Network hybrid framework is now my focus. The database we use is IAM Handwriting Databse.

In this project I worked on:
  • Building baseline GMM-HMM system
  • Building RNN-HMM system
  • Visualization tool to compare with Evernote
  • RNN based online offline handwriting conversion to deal with data shortage problem

Advisors: Frank Soong, Lijuan Wang

HMM-DNN Audio Visual Speech Recognition

In this project, we try to use combined audio visual feature to improve the accuracy of speech recognition. The data set we use is from the MOOC videos of Cousera. Based on context dependent DNN-HMM structure, we investigated the methods of combination, and compared the results of single feature or combined feature.

I am fully involved of this project.

Advisors: Frank Soong, Lijuan Wang

Grow Search

Inspired by Beyond Search project from Microsoft Research Cambridge, we create a prototype software to visualize the search results using a "growing tree". Each keyword the user type in can become a tree, and the search results are represented as the leaves. As time goes by, the tree will grow. By typing more keywords, the user can finally get a forest about a specific area. We believe it is a clear and also fun way to do the searching on the Internet.

A video introduction can be watched here.

In this project I play as the product manager of the developing team.

Parallel Corpus Extraction From Bilingual Govermental Websites

There are lots of valuable resource on Internet which can provide with cross languages and cross areas parallel corpus. Some earlier methods are developed to do this mining work. However, they often use one feature only in the mining process. We use multiple reasonable features of parallel pages to acquire parallel corpus. At last, we also add a SVM classifier which utilize all the features to do the mining work. Surely, it achieve a significant improvement than earlier methods. The evaluation is based on massive manually annotated pairs and our method achieves precision rate of 95% and recall rate of 99%.

This work has been accepted by Journal of Computational Information Systems (JCIS), 2012. [PDF]

I am in charge of parallel sentences extraction part.

Advisor: Dequan Zheng

HIT ACM Online Judge

To give a better experience of the programming fans at HIT, we re-design the Online Judge System, both in front-end and back-end. In front-end we give a cleaner and more elegant user interface, and fix several places which are not very user friendly. In back-end we use a new framework to manage different layers (actions and views) of code. This improves the maintainability compared to the previous "naked" PHP code.

I am mainly in charge of the front-end developing, and some back-end developing.

Copyright © Xiong Zhang 2016. All rights reserved.