PadmanabanVarma

Introduction

Although camera technology is steadily advancing year-after-year, one of the problems that continues to plague photography is that of low-light noise. With the widespread adoption of smartphones, everyone today has access to a high quality camera in their pocket, and with it they document their lives at all times of the day. However, once the scene goes dark all of these images begin to be dominated by noise, specifically poisson shot-noise.

Initial efforts started with simple denoising techniques such as Wiener filtering \cite{wiener}, but they only work well if the underlying image is smooth. Another set of denoising techniques relied on looking at the Fourier or wavelet domain \cite{wavelet}. These were better at preserving images while denoising the images. However, they are usually very complex and computationally expensive. Moreover, all described approaches focus only on removing Gaussian noise, which is nor what dominates low-light images. The ICA method \cite{ica}, on the other hand, is able to deal with non-Gaussian data. Even though it works well with low-light images, the technique requires many image frames of the same scene, which isn't possible for denoising pre-captured images.

We attempt to solve this problem by learning a linear-regression model which, given a target area to denoise and local image information around the target, can correctly predict the true value of the pixels. And applied repeatedly, our method can then denoise an entire input image. [1]

[[File:OculusPipeline.pnlink titleg | 900px | center| thumb | caption | Fig. 1 – Streaming and Augmenting Stereo Camera Images-Pipeline ]]

References

dkfjsf