LATEST PROJECTS

Built a CycleGAN-based dark image enhancement model using Keras. Extracted content loss by VGG to preserve input texture. For deblurring and sharpening, calculated edge histogram loss as the adaptive weight of illumination loss.
Detected facial landmarks, used skin detector and edge detector for facial landmarks correction. Extracted geometrical features by landmarks, and designed a frame-based facial expression recognition system.
Trained SVM, random forest and MLP for facial expression recognition, got the accuracy of 82%.

Designed and implemented a Full Stack, RESTful Yelp-like campground rating app that supported user log in, managing

profile page, and posting review and comments.
Used Node.js and Express.js to create multiple routes that rendered ejs templates.
Stored user, campgrounds and comments data and handled their association by MongoDB.

Managed an Agile project to improve the user interface for a Ruby on Rails-based shuttle call application.
Designed a relational database including vehicle, customer, trip and location, made the E-R diagram and normalized

database into Boyce-Codd normal form.
Added user profile page using HTML/CSS, attached PostgreSQL to this page.
Utilized ArcGIS to add interactive web map to facilitate users’ route planning.

Utilized Stanford CoreNLP to obtain sentiment scores for top 25 news headlines in the last 16 years, trained Random Forest,

Naïve Bayes and SVM models to predict stock market based on sentiment scores, got the best accuracy of 53%.
Converted words in news to word vector by Word2vec, trained a LSTM model to predict stock market based on word vectors

in the previous time, got an accuracy of 53%. Then, clustered news by 30 topics LDA and improved accuracy by 4%.