Zoom Link: https://hku.zoom.us/j/9174230511?omn=91474809260
Abstract
In recent years, deep neural networks (DNNs) have achieved remarkable success across a wide range of applications, especially for low-level image processing. However, many of these approaches rely on complicated and deeply stacked architectures to attain high performance, which lead to significant challenges for deployment on edge devices with limited computing resources. Therefore, research has been motivated to design efficient pipelines to enable practical and resource-friendly AI applications without compromising accuracy. We explore the efficient architecture design of low-level image processing. To avoid the floating-point operations introduced by convolutional neural networks (CNNs) and use as little space as possible, we design various pure Lookup Tables (LUTs) framework for three widely applied tasks, covering the single image super resolution, color enhancement, and 3D reconstruction.
Speaker
Binxiao Huang
Department of Electrical and Electronic Engineering
The University of Hong Kong
Speaker’s Biography
Binxiao Huang received his B.Eng. Degree and master’s degree from Beihang University. He is currently pursuing a PhD degree in the Department of Electrical and Electronic Engineering, The University of Hong Kong. His research focuses on efficient deep learning methods for image processing and 3D reconstruction.
All are welcome!