UROP Proceedings 2020-21

School of Engineering Department of Computer Science and Engineering 118 Deep Video Super-Resolution Supervisor: CHEN Qifeng / CSE Student: WU Yi / COSC Course: UROP1100, Summer Text image super resolution is an important sub-field of Super Resolution, which aims to recover the realistic features of a text image from a low-resolution text image. Recovering the realistic features from a lowresolution image is challenging and it has drawn much research interest of the computer vision community. In our project, we self-learned some basic knowledge of deep learning from cs 231n and collected thousands of screenshots as a dataset, then use this dataset to train two open-source deep learning models. The training result shows these models show surprising performance on highly blurred screenshot images, and we plan to improve these models to achieve higher performance. Deep Video Super-Resolution Supervisor: CHEN Qifeng / CSE Student: XIAO Ziruo / SENG Course: UROP1100, Summer In the scene of text image got zoomed or compressed, low-resolution often results in the loss of text information. Our group focus on the super-resolution of text image, trying to get the text clear enough to be recognized. We collected data, run them on some existed models, trying to figure out their shortcomings and hopefully build one better for Chinese text in the future. As a year 1 student introduced to computer vision and deep learning for the first time, I put most of my effort in learning and practicing the basic knowledge of the convolutional neural network in this summer. This report will show what I have learned in this 3 months, the project progress and my future plan. Deep Video Super-Resolution Supervisor: CHEN Qifeng / CSE Student: YAO Chongchong / DSCT Course: UROP1100, Summer As compression technology is commonly used, compressed text pictures may be blurred and difficult to read. In such cases, it would be helpful to restore their original sharpness, which is where super-resolution technology comes in. This project aims to learn and apply super-resolution models to computer-generated text images to increase their resolution. To achieve our goal, our team managed to collect naïve data and processed them to generate training and testing samples. We first tested our data on several pre-trained models and compared the results. In the end, we successfully trained several models and got satisfying sharpened outputs. Overall, this study produced some practical results for computer-generated text image super-resolution and also improved our knowledge and technique on deep learning.