A New Method Of Poisson Gaussian Signal-Dependent Noise Parameter Estimation Based On Weakly Textured Image Blocks Selection

As an important carrier of information recording and transmission in modern society,images are inevitably affected by noise in the process of collection,storage,and transmission.Complementary metal oxide semiconductor is one of the commonly used image sensors in modern digital imaging systems.A complex noise is generated during the imaging process.This noise is not simply additive white Gaussian noise,but a signal-dependent noise that is highly correlated with the pixel level.The noise level is an important parameter of image processing and image optimization algorithms such as image denoising,image compression,and image stitching.Therefore,it is of great significance to accurately estimate the CMOS noise level.At present,most estimates of signal-dependent noise are based on weakly textured image blocks.Because the pixel levels in the local weakly textured area of the noise image are basically the same,signal-dependent noise can usually be approximated as additive noise,so the key to estimating the noise level of signal dependent noise lies in the selection of weak texture image blocks.The accuracy of the result of the weak texture image block selection has a direct impact on the noise parameter estimation result,and is one of the key factors that determine the estimation effect.In order to improve the accuracy of signal-dependent noise estimation,this thesis starts with the selection of weakly textured image blocks,and strives to find out the image blocks with weaker texture more accurately.In the process of selecting weakly textured image blocks,the definition of the texture intensity of the image block and the determination of the block selection threshold are the two most critical steps to determine the accuracy of the block selection result.Some existing methods of texture intensity definition and threshold determination are still insufficient.For the definition of image block texture intensity,this paper proposes two image block texture intensity definition methods,which are based on local binary cyclic jump and texture information statistics.For the definition of image block texture intensity,this paper proposes two image block texture intensity definition methods,which are based on local binary cyclic jump and texture information statistics.In order to extract the texture information in all directions as much as possible,the local binary cyclic jump is used to integrate the texture intensity information in all directions around a single pixel.Then from the point and surface,the texture intensity information around each pixel is averaged and integrated to obtain the texture intensity of the image block.Different from the former,the method for defining the texture intensity of an image block based on the statistics of texture change information starts from a small 3×3 image block.In order to count the texture information in each direction,this method calculates the texture information between two adjacent pixels from the horizontal direction,vertical direction,main diagonal direction and sub-diagonal direction,a 3×3 Small image blocks can get 20 sets of texture information,which are accumulated and integrated as the texture information of small image blocks.Finally,the texture information of the N×N image block is calculated as the average value of the texture information of all 3×3 small image blocks in the block.For the determination of the block selection threshold,this paper adopts two methods.First,one way is to design experiments to determine based on experimental results.Another method is to first derive the mathematical statistical law of the texture intensity of flat image blocks affected by noise.This law explains the mapping relationship between noise level and texture intensity after a flat area is affected by noise.This relationship can calculate the texture intensity of the flat area affected by noise as the block selection threshold,but the noise level needs to be known in the calculation,so a method for iterative selection of weak texture image blocks is designed.After selecting a suitable weak texture image block,for each weakly textured image block,its pixel level and noise variance are estimated as a set of sample pairs,and finally these sample pairs are fitted to obtain noise parameters.Experimental results show that the two signal-dependent noise estimation algorithms proposed in this paper can obtain estimation results with smaller mean square error compared with the existing methods.