Noise Suppressing In Low-Light Digital Camera Processing Pipeline

Recent progress on digital camera technology has had extraordinary impact on numerous areas,such as Aerospace,information security,biological engineering,artificial intelligence and etc.A signal processor named as 'digital camera pipeline' embedded in digital camera generates a digital full color image from raw sensor data captured by camera sensor.A dig-ital camera pipeline generally comprises sensor data readout,denoising,demosaicking,color correction,white balance,gamma correction and etc.Each of the process is crucial for gener-ating high quality digital images.Under ideal imaging conditions,a decent image can be pro-duced by the baseline processing that consist of demosaicking,color correction,white balance and gamma correction.This baseline processing is the basis of all digital camera processing pipeline.However,most images are taken under nonideal lighting,with imperfect hardware,and by a nonexpert photographer.These less ideal imaging conditions have adversary effects on the output image quality.Hence advanced processing tasks are needed to increase robust-ness against them.Under low light scenario,noise remains a serious problem in image sensor data,contaminating image details and tectures as well as increasing the difficulty to the image processsing steps like demosaicking.Therefore,noise suppression techniques are required in the digital camera processing pipeline for recovering high quality images.In this dissertation,we propose several methods to improve the image quality in low light scene.The main contributions in this dissertation are organized as follows:1.Binning reduces the impact of read noise on the combined signal even if the individual pixel values are small.However,the noise performance improvements come at the price of spatial resolution loss.A new pixel binning design for color image sensors aimed at minimiz-ing the resolution loss while improving noise performance.The proposed algorithm combines neighboring pixels to yield superpixels arranged in a square sampling lattice.Compared to ex-isting binning methods,the proposed binning has considerably less spatial and color artifacts and better noise performance.2.A binning based single-shot HDR imaging scheme has been proposed for reproducing a greater dynamic range luminosity of low light image suffering under exposure or over exposure.Two low dynamic range(LDR)images have been generated via the binning scheme proposed in this paper and then combined to produce a HDR image.The combined HDR image effectively improves the performance,with little noticeable influence of noise and saturation.3.This paper investigates discrepancies between the model of image sensor noise com-monly used in image denoising algorithms and the distribution of actual sensor noise acquired by real.Existing studies overwhelmingly focused on the relationship between the pixel intensity and noise variance,but there has been little emphasis on the actual distribution of the noise.As the tail behavior of the noise distribution greatly influences denoising performance,we provide detailed analysis of this.Moreover,most modern image denoising techniques incorporate linear and nonlinear transformations that give rise to energy compaction and sparse signal represen-tation.Therefore we focus our investigation on the accuracy of noise model in the transform domain rather than the noise distribution in the pixel intensity domain.4.A new Gaussian mixture noise model is proposed to correct the mismatch of tail behav-ior.Based on the fact that noise model mismatch results in image denoising that undersmoothes real sensor data,we propose a mixture of Gaussian denoising method to remove the denoising artifacts without affecting image details,such as edge and textures.The“Poisson mixture image denoising”problem can be solved by "a mixture of Poisson image denoising.".Exper-iments with real sensor data verify that denoising for real image sensor data is indeed improved by this new technique.5.A wavelet-based Gaussian scale mixture(GSM)joint demosaicking and denoising method has been proposed in this paper.The wavelet coefficients of the proposed method cor-responding the luminance and chrominance components are reconstructed using Bayesian min-imum mean square error(MMSE)estimation.The proposed wavelet-GSM technique exploits the correlation of neighboring wavelets coefficients and makes the denoising in demosaicking explicit.As a result,the proposed demosaicking method effectively suppresses the zippering artifacts and improves the robustness to noise than the state-of-the-arts.