Simulating Images with Different Camera Settings

MPhil Thesis Defence


Title: "Simulating Images with Different Camera Settings"

By

Mr. Hao OUYANG


Abstract

We introduce a camera simulator to synthesize raw sensor data under 
different camera settings, including exposure time, ISO, and aperture. The 
simulator consists of three components:  an exposure module relying on the 
principle of modern lens designs, a noise module utilizing deep 
convolutional denoising networks and noise level functions, and an 
aperture model using a new adaptive attention module. Through the proposed 
pipeline, we can correct the luminance level, adapt the noise, and 
synthesize the defocus blur. We collected a dataset of more than ten 
thousand raw images of 450 diverse scenes with different camera settings 
using two cameras. The dataset can not only facilitate the training of the 
simulator but also benefit other tasks such as training image descriptors. 
Quantitative comparisons and qualitative results demonstrate that our 
approach outperforms relevant baselines in raw data simulation. 
Furthermore, our camera simulator can enable multiple applications, 
including large-aperture enhancement, HDR, and training auto exposure 
mode. Our work represents the first effort to fully simulate a camera 
sensor's behavior, leveraging both the power of conventional raw sensor 
characteristics and the potential of data-driven deep learning.


Date:  			Monday, 17 August 2020

Time:			2:30pm - 4:30pm

Zoom meeting:		https://hkust.zoom.us/j/91709394400

Committee Members:	Dr. Qifeng Chen (Supervisor)
  			Dr. Xiaojuan Ma (Chairperson)
 			Prof. Chiew-Lan Tai


**** ALL are Welcome ****