저조도 환경에서 휴대폰 카메라에 사용될 프레임 메모리를 사용하지 않는 실시간 디지털 이미지 안정화 기법

Abstract

Due to its advantages of low-cost and lower power consumption, the complementary metal– oxide– semiconductor image sensor (CIS) is widely used in low-cost cellular phone cameras. However, with chip size becoming smaller but spatial resolution increasing, the image sensor of a camera cannot usually correctly capture images in low ambient light conditions, especially cellular phone cameras that use a fixed aperture and cannot adjust for the amount of incident light. Therefore, users have to extend exposure time to collect more light and enhance the signal-to-noise ratio (SNR). However, longer exposure results in motion blur due to inevitable hand movement or object motion. If exposure time is not extended, but the rule of thumb of short exposure time under a safe shutter speed is followed, users can avoid motion blur; however, this method usually generates under-exposed images. Optical image stabilization (OIS) methods utilize inertial sensors to measure camera motion and then adjust the lens or shift the sensor accordingly based on the motion magnitude and direction. OIS can achieve good performance; however, it requires additional apparatuses and occupies more space. Digital image stabilization (DIS) tries to generate acceptable output images by restoring the blurred image, merging several varied exposed images, or correcting the under-exposed image. Many conventional DIS methods achieve good performance; however, most of them can only be implemented by software with a large delay. To perform image stabilization in real-time, two hardware-oriented DIS are proposed in this dissertation. The first method post-processes an image captured with a safe shutter speed using an adaptive denoising filter and a global color correction algorithm. The proposed algorithm can transfer the normal brightness of an image previewed in a long exposure time to the under-exposed snapshot, making a bright and crisp image with low noise. The second method proposes a novel imaging method using no-flash and flash image pairs, the former usually have too much noise, while the latter have a high SNR but look unnatural. By transferring the natural tone of the former to the latter, the resulting image can have a high SNR and maintain a natural appearance. In order to implement the two proposed methods in real-time, this dissertation uses two preview images to estimate the tonal transformation function in advance, and then uses this function to adjust the tone of the captured image in real-time. The two proposed methods do not require frame memory and can be implemented with hardware and in real-time.