UROP Proceedings 2020-21

School of Engineering Department of Computer Science and Engineering 116 Deep Video Super-Resolution Supervisor: CHEN Qifeng / CSE Student: SUN Yushi / COSC Course: UROP1100, Fall In the field of photography, it's always desirable to remove the polarized reflection from the image, since the reflection actually influence the visual quality of a photo. However, this is not an easy task, since it's hard for machine, even human, to tell which part belongs to the reflection that we want to remove and which part belongs to the original image. In this project, we want to identify the potential of employing super resolution models to perform polarized reflection removal task. We will first build up some existing super resolution models and then test them on the Polarized Reflection dataset provided by [1]. The model structure and the experiment result will be demonstrated. Since this is just a progress report, we will also introduce our future plan at the end of this report. [1] C. Lei, X. Huang, M. Zhang, Q. Yan, W. Sun, and Q. Chen, “Polarized reflection removal with perfect alignment in the wild,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 1750–1758. Deep Video Super-Resolution Supervisor: CHEN Qifeng / CSE Student: ZHOU Ji / CPEG Course: UROP1100, Fall The project focuses on applying super resolution models on multiple data sets. With the testing and training of data sets based on polarized reflection removal and multiscopic vision, we expect to improve the original model and increase the accuracy of the model. In this project, I intended to use SRGAN as the SR model. Unfortunately, due to the lack of knowledge about machine learning and computer vision, and the tight time schedule, the final outcomes is not as well as we expected. This report will introduce related resources such as the applied model and the data sets used with a short explanation and show the process of the project in this term. Deep Video Super-Resolution Supervisor: CHEN Qifeng / CSE Student: JIANG Xudong / COSC Course: UROP1100, Spring With the development of deep learning, various computer vision tasks have made significant progress. One of the important computer vision tasks, image and video superresolution, also experienced remarkable improvement. Various deep learning models have been proposed to solve the video superresolution problem. However, there is a popular video format, GIF, which most existing models cannot perform well. This result from the artifacts caused by severe color quantization and color dithering. In this project we review recent deep learning techniques about video super resolution and extend the application to GIF format videos. We obtained some perceptually more satisfying results in the experiments, although the result in objective evaluation metrics is still in need of improvement.