Xiaoshuai Zhang

UC San Diego email: zхs АТ uсsd DОТ еdu

Hi, I'm Xiaoshuai (Jet) Zhang (张孝帅, IPA: /ʈ͡ʂɑŋ ɕiau ʂuai/), a first-year Ph.D. Student supervised by Prof. Hao Su at the Dept. of Computer Science and Engineering, UC San Diego (CSE, UCSD). My general research interests cover Computer Vision and Machine Learning. I received my B.S. in Machine Intelligence from Peking University, advised by Professor Jiaying Liu. I also worked with Haoqiang Fan at Face++.

For leisure, I'm interested in photography, linguistics, phonology and all modes of transport. Cities: Skylines is my favorite video game.

To English speakers: Shaun-shy is a good approximation for my first name :).



1. Dynamically Unfolding Recurrent Restorer: A Moving Endpoint Control Method for Image Restoration

Xiaoshuai Zhang*, Yiping Lu*, Jiaying Liu and Bin Dong.

Accepted by the International Conference on Learning Representations (ICLR), 2019. [OpenReview] [Project Page] [Code]

2. AdaMatting: Integrated Trimap Adaptation and Alpha Estimation

Shaofan Cai*, Xiaoshuai Zhang*, Haibin Huang, Jiangyu Liu, Biao Leng, Jiaying Liu, Jue Wang, and Haoqiang Fan.

Submitted to some CV conference, 2019. [arXiv (to be released)]

3. Bridging the Gap Between Computational Photography and Visual Recognition

Submitted to some journal, 2019. [arXiv]


1. DMCNN: Dual-Domain Multi-scale Convolutional Neural Network for Compression Artifacts Removal

Xiaoshuai Zhang, Wenhan Yang, Yueyu Hu and Jiaying Liu.

Accepted by the IEEE International Conference on Image Processing (ICIP), 2018 as an oral presentation. [Paper] [arXiv] [Project Page] [code]

More publications coming soon ......



1. Dynamically Unfolding Recurrent Restorer (DURR)

We propose a new control framework to restore images corrupted by different degradation levels. The proposed control problem contains a restoration dynamics modeled by an RNN, and the terminal time is determined by a policy network. The model is called DURR. Experiments show that DURR achieves state-of-the-art performances on blind image denoising and JPEG image deblocking. Furthermore, DURR is able to generalize to damage levels not included in the training stage.

My Homepage
2. JPEG Image Deblocking: DMCNN

An end-to-end CNN with large receptive fields exploiting dual-domain multi-scale features. The proposed DMCNN can handle multiple kinds of JPEG compression artifacts, especially banding effects effectively. Experiments show that DMCNN sets a new state-of-the-art for the task of JPEG artifact removal.

My Homepage
3. HEVC Post Filter

We are trying to develop a convolutional neural network to work as an HEVC post-process filter to remove blocking artifacts and ringing artifacts, especially at low bit-rates.

My Homepage
4. Online Badminton

A javascript rewrite of the Flash mini-game Stick Badminton 2. The project report is here (Chinese).


My Homepage
1. Visualization of Beijing House Data

A visualization of Beijing house data, includes house prices, sizes, locations and so on. The data are from Lianjia Beijing.

My Homepage
2. Facial Master

A hackathon project like Apple's animoji. We use a webcam to extract face landmarks and analyze people's expression. Then map these facial expressions to a series of pre-defined open-source face models in the real-time.

My Homepage
3. Chrys: An AI for Multi-Player Pacman

We develop an AI for the game of Pacman on the Botzone (An online AI judge platform) utilizing mainly the Monte Carlo Tree Search (MCTS) and other evaluation techniques. Chrys ranks 1st on the platform till now. You can view a sample game here. (My Botzone IDs are @BuriedJet and @蒓凊尐囗�?)


My Homepage
1. Ajax: An AI for Reversi

We develop an AI for the game of Reversi on the Botzone (An online AI judge platform). Ajax is based on decision tree searching with alpha-beta pruning and other search techniques (eg. Iterative Deepening and PVS). Featuring a sophisticated evaluation function, Ajax ranked 1st at its birth and maintains the top ten till now. The codes can be found here. You can play with Ajax online here.