Single-image Signal-dependent Noise Parameter Estimation Method Based On Relative Density Peak Clustering

With the continuous development of microelectronics science and technology,the pixel size of image sensors is becoming smaller and smaller,and signal-dependent photonic noise has become the main source of noise in images,leading to degradation of image quality,therefore,the removal of signal-dependent noise plays a key role in improving image quality.Most of the existing methods for estimating signal-dependent noise parameters are based on weakly textured image blocks,and the key to this method lies in the selection of weakly textured image blocks.Designing a more accurate weakly textured image block selection method can improve the accuracy of noise parameter estimation and thus improve the execution of subsequent image denoising work.The existing methods for selecting weakly textured image blocks suffer from the problems of single image representation information and the setting of the block selection threshold being influenced by human experience,leading to the inaccurate estimation results of the noise parameters.Therefore,we propose a new single-image signal-dependent noise parameter estimation method based on relative density peak clustering,which selects weakly textured image blocks by extracting multiple features of the image for clustering,overcoming the above problems.However,the traditional density peak clustering algorithm is based on absolute density values when calculating local density,which makes it difficult to select clustering centers of sparse clusters in datasets with uneven density distribution,as they are often ignored or merged into other clusters,leading to inaccurate clustering results.Therefore,we propose the relative density peak clustering(RDPC)algorithm,which eliminates the influence of density distribution differences between clusters and allows sparse clusters to be selected accurately.The main work of this paper is:Firstly,in response to the problem that the traditional density peak clustering algorithm is difficult to select sparse clustering centers in datasets with uneven density distribution,resulting in inaccurate clustering results,a relative density peak clustering algorithm is proposed.First of all,information entropy is introduced,and the cutoff distance is selected adaptively through the calculation of information entropy,avoiding the influence of empirically setting of the cutoff distance on the clustering results.A relative density definition is then designed to replace the traditional definition of absolute local density,considering the local structure of the data,eliminating the influence of clustering results by differences in distribution density between clusters,and allowing even sparse density clusters to be accurately selected.The proposed RDPC algorithm is compared with two classical clustering algorithms: k-means clustering algorithm and density-based spatial clustering of applications with noise,the original clustering algorithm and three improved clustering algorithms:density-normalized density peak clustering,relative density-optimized density peak clustering and domain adaptive density clustering on nine commonly used synthetic datasets and eight real datasets,and the performance of the clustering algorithm is measured by three criteria: F-score,normalized mutual information and adjusted rand index.The experimental results show that the mean values of F-score,adjusted rand index and normalized mutual information of the proposed algorithm are 14%,30% and 28% higher than those of other algorithms,respectively,and the clustering results are more accurate on both datasets with even density distribution and those with uneven density distribution.Secondly,to address the problems of single image information representation and the setting of block selection thresholds affected by human experience in existing weakly textured image block selection methods,a method based on relative density peak clustering is designed to select weakly textured image blocks and then perform signal-dependent noise parameter estimation,and a singleimage signal-dependent noise parameter estimation method based on relative density peak clustering is proposed for the first time.The proposed method is experimented on two commonly used noise models: Poisson-Gaussian noise model and Gaussian-Gaussian noise model,respectively.The accuracy of the noise parameter estimation is evaluated with Mean Square Error(MSE).The proposed single-image signal-dependent noise parameter estimation method is compared with the method based on image gradient matrix,the method based on image histogram,the method based on image grey entropy and the method based on weight and shape constrained.The experimental results show that the proposed method achieves smaller MSE on both noise models compared to the other methods,indicating that the method is effective in improving the accuracy of signal-dependent noise parameter estimation.