Signal-dependent noise parameter estimation method based on improved density peak clustering

Abstract

The present invention discloses a signal-dependent noise parameter estimation method based on improved density peak clustering. Samples are extracted from a noisy image through a sliding window, and the mean, entropy and gradient are calculated as feature data. Then the samples are input into a clustering algorithm for clustering to distinguish weak texture samples from strong texture samples. In the clustering process, the concept of relative density is introduced. The comparison range is divided by the distance between the data point and the surrounding data. The relative density of the data points in each comparison range is calculated, and finally the data points with high relative density are selected as the clustering center, which solves the problem that the traditional DPC algorithm often ignores the sparse cluster center when clustering data sets with uneven density, thereby affecting the clustering accuracy. According to the clustering results, pixel level estimation and noise estimation are performed on samples with cluster labels as weak textures. Finally, the pixel value-noise variance estimation pair is fitted by the least squares method to obtain the noise parameter estimation value of the original image, thereby realizing the preparation work for denoising.

Description

Technical Field

[0001] The invention belongs to the technical field of image noise signal processing, and in particular relates to a signal-dependent noise parameter estimation method based on improved density peak clustering. Background Art

[0002] Images are important carriers for information recording and transmission in modern society, but images will inevitably be affected by noise during the process of acquisition, storage, and transmission, thereby reducing the quality of the image. The noise level is an important parameter of image processing optimization algorithms such as image denoising, image compression, and image stitching. Therefore, accurately estimating the noise level is of great significance for image denoising. With the development of CMOS image sensors, in order to reduce the impact of noise, some manufacturers directly embed noise reduction modules in image sensor chips, which can effectively suppress signal-independent noise. Signal-dependent noise components are becoming the main noise source of CMOS image sensors. At present, most of the estimation of signal-dependent noise is based on weak texture image blocks. The difficulty of this method lies in the selection of weak texture image blocks, and then estimating a pair of pixel values ​​and noise variance for each image block, and finally fitting the noise parameters. Weak texture image blocks can be obtained using clustering algorithms.

[0003] Clustering is the process of grouping different objects according to their similarities, and objects with high similarity are grouped into the same group. Cluster analysis, as the main content of unsupervised learning, has always been an important part of machine learning, pattern recognition, data mining and other fields. At present, many clustering algorithms have emerged, which can be roughly divided into partition clustering method, hierarchical clustering method, density clustering method, model clustering method, grid clustering method, etc.

[0004] In 2014, Rodriguez and Laio proposed a density peak clustering algorithm (DPC) based on density and distance in Science. The algorithm can quickly search and find density peaks, then determine the cluster center, and assign the remaining points to the corresponding clusters. The DPC algorithm has the advantages of few parameter requirements and non-iterativeness, and can effectively detect clusters of arbitrary shapes from large-scale data sets with low computational complexity. However, the existing density peak clustering algorithm tends to select the cluster center of dense areas and often ignores the clusters of sparse areas. For data sets with large density differences, the DPC algorithm often determines the density points in sparse areas as outliers or misplaces them to adjacent dense clusters, thereby ignoring the cluster center of the sparse cluster, resulting in inaccurate clustering results. Some existing improved clustering algorithms based on DPC often improve the clustering effect by introducing k-nearest neighbors (KNN), but the parameter k is a numerical value set subjectively based on past experience, and the size of the k value often has a greater impact on the clustering results, which further affects the accuracy of clustering. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes a signal-dependent noise parameter estimation method based on improved density peak clustering, introduces the definition of relative density, determines the clustering center according to the distribution of image features, does not require manual definition of parameter values, improves clustering accuracy, and thus achieves accurate estimation of noise parameters.

[0006] The signal-dependent noise parameter estimation method based on improved density peak clustering specifically includes the following steps:

[0007] Step 1: Use a sliding window to extract several samples of the same size from the original image containing noise. Calculate the mean, entropy and gradient of each sample as the feature data of the sample. The feature data of all samples form a feature data set, which is then normalized.

[0008] Step 2: Calculate the following parameters for the feature data set X after normalization in step 1:

[0009] ①Feature data x i The Euclidean distance d between other feature data ij :

[0010] d ij =||x i -x j ||

[0011] ②Feature data x i The local density ρ i:

[0012]

[0013] ③Feature data x i The distance δ to the nearest higher density feature data i :

[0014]

[0015] ④Characteristic data x i The distance θ to the nearest feature with lower density i :

[0016]

[0017] Among them, x i 、x j ∈X,d c is the cutoff distance, which is the value of 1% to 2% after the distances between all feature data are arranged in ascending order.

[0018] Step 3: Divide the feature data x according to the calculation results of step 2 i Relative density comparison range η i And calculate its relative density ρ i ':

[0019] η i =max(θ i ,δ i )

[0020]

[0021] Among them, N i is the relative density comparison range η i The number of feature data in . Represents feature data x j The weight of the distance feature data x i The closer, the better for the feature data x i The greater the relative density contribution, the greater the weight. Define the characteristic data x i The relative density is the ratio of its local density to the cumulative average of the local density of the surrounding feature data.

[0022] Step 4: Use the feature data x obtained in step 3 i The relative density ρ i 'Replace the local density ρ i , recalculate the feature data x i The distance δ to the nearest higher density feature i Then calculate the characteristic data x iThe clustering value γ i :

[0023] γ i =ρ i '·δ i

[0024] The clustering value γ is calculated i Sort in descending order, select the two feature data with the highest clustering values ​​as cluster centers, assign other feature data to clusters with closer cluster centers, and then output the cluster label corresponding to each feature data to complete the clustering process.

[0025] Step 5: Calculate the estimated pixel level of the weak texture sample based on the clustering results of step 4 and the noise variance

[0026] wk=1~B_weak_n

[0027] wk∈[1,B_weak_n]

[0028] Where N is the side length of the sample, wk represents the index of the weak texture sample, B_weak_n is the number of weak texture samples, I wk (m,n) represents the weak texture sample I wk The pixel value of the mth row and nth column. is the minimum variance direction vector of the weak texture sample, and the covariance matrix C P The smallest eigenvalue of is related to:

[0029]

[0030] Step 6: Configure the noise in the original image as a Poisson-Gaussian noise model, then the overall noise method σ at the (p, q) position in the original image 2 (p,q) is:

[0031] σ 2 (p,q)=ax(p,q)+b

[0032] Where x(p,q) represents the noise-free pixel value at position (p,q) in the original image. a and b are noise parameters. Weak texture sample I is fitted using the least squares method. wk The pixel value-noise variance estimate pair Get the estimated value of the original image noise parameter and

[0033] The present invention has the following beneficial effects:

[0034] 1. For a dataset with uneven density distribution, such as characteristic data of signal-related noise of CMOS image sensors, this application defines a relative density in the process of clustering samples. The relative density is calculated based on the data distribution characteristics of the dataset itself, so that the center point of a sparse cluster can also have a larger density value, thereby preventing the density value of a sparse cluster from always being lower than the density value of a dense cluster in a dataset with large density differences, and avoiding treating a sparse cluster as an outlier or misassigning it to other high-density clusters.

[0035] 2. The comparison area is divided according to the data distribution characteristics of the data set, without introducing subjective parameter values, so that different data sets have different relative density comparison ranges adaptively, avoiding the influence of human experience judgment on the experimental results.

[0036] 3. The introduction of relative density improves the accuracy of clustering results, thereby selecting more accurate weak texture areas, and then estimating the noise of the entire image, improving the estimation accuracy, and facilitating subsequent image denoising. DETAILED DESCRIPTION

[0037] The present invention is further explained below. The original image used in this embodiment comes from an image captured by a CMOS image sensor, and the noise in the original image is a signal-dependent noise component generated by a noise reduction module of the CMOS image sensor during the acquisition process.

[0038] The signal-dependent noise parameter estimation method based on improved density peak clustering specifically includes the following steps:

[0039] Step 1: Use a 16*16 sliding window to extract n 16*16 samples from the original image containing noise, sliding one pixel at a time from top to bottom and from left to right. Calculate the mean, entropy and gradient of each sample as the feature data of the sample:

[0040]

[0041]

[0042]

[0043] Among them, gray represents the size of the gray value, and p(gray) represents the probability that the gray value is gray. W and H are the width and height of the sample respectively, and p w,h Represents the pixel value at position (w,h) in the sample.

[0044] That is, the characteristic data x of each sample i’It has three properties: mean, entropy and gradient, x i' ={x i1 ,x i2 ,x i3}. The feature data of all samples form a feature data set and are normalized.

[0045] Step 2: Calculate the following parameters for the feature data set X after normalization in step 1:

[0046] ①Feature data x i The Euclidean distance d between other feature data ij :

[0047] d ij =||x i -x j ||

[0048] ②Feature data x i The local density ρ i :

[0049]

[0050] ③Feature data x i The distance δ to the nearest higher density feature i :

[0051]

[0052] ④Characteristic data x i The distance θ to the nearest feature with lower density i :

[0053]

[0054] Among them, x i 、x j ∈X,d c is the cutoff distance, which is the value of 1% to 2% after all the feature data are arranged in ascending order. i When the density of is the global minimum, θ i =δ i .

[0055] Step 3: Divide the feature data x according to the calculation results of step 2 i Relative density comparison range η i And calculate its relative density ρ i ':

[0056] η i =max(θ i ,δi )

[0057] When η i =δ i When the characteristic data x i A point that is surrounded by feature data with lower density is more likely to become a center point.

[0058] When η i =θ i When the characteristic data x i Surrounded by feature data with higher density, the probability of becoming the center point is small.

[0059]

[0060] Among them, N i is the relative density comparison range η i The number of feature data in . Represents feature data x j The weight of the distance feature data x i The closer, the better for the feature data x i The greater the relative density contribution, the greater the weight. Define the characteristic data x i The relative density is the ratio of its local density to the cumulative average of the local density of the surrounding feature data. In this way, the density value calculated for each point is the relative size of the density of the point and its surrounding points, rather than the absolute size, which better considers the local structure of the cluster. Even the cluster center of a sparse cluster can obtain a larger density value, thereby better identifying the sparse cluster.

[0061] Step 4: Use the feature data x obtained in step 3 i The relative density ρ i 'Replace the local density ρ i , recalculate the feature data x i The distance δ to the nearest higher density feature i The traditional DPC algorithm needs to draw a decision diagram based on ρ and δ, and then manually select the points with larger ρ and δ as the cluster centers. This method has a large proportion of subjective factors, and there is also a situation where the cluster center cannot be easily selected by the naked eye, which will directly lead to a domino effect in the subsequent allocation steps and affect the final clustering results. Therefore, this application chooses to calculate the product of relative density and distance as the feature data x i The clustering value γ i :

[0062] γ i =ρ i '·δ i

[0063] The clustering value γ is calculatedi Sort in descending order, select the two feature data with the highest clustering value as the cluster center, assign other feature data to the cluster where the cluster center is closer, and then output the cluster label corresponding to each feature data. The cluster labels are divided into weak texture samples and strong texture samples. After completing the clustering process, the weak texture sample can be found.

[0064] Step 5: Calculate the estimated pixel level of the weak texture sample based on the clustering results of step 4 and the noise variance

[0065] wk=1~B_weak_n

[0066] wk∈[1,B_weak_n]

[0067] Where N is the side length of each sample, in this embodiment N = 16, wk represents the index of the weak texture sample, B_weak_n is the number of weak texture samples, I wk (m,n) represents the weak texture sample I wk The pixel value of the mth row and nth column. is the minimum variance direction vector of the weak texture sample, and the covariance matrix C P The smallest eigenvalue of is related to:

[0068]

[0069] Step 6: Configure the noise in the original image as a Poisson-Gaussian noise model, then the overall noise method σ at the (p, q) position in the original image 2 (p,q) is:

[0070] σ 2 (p,q)=ax(p,q)+b

[0071] Where x(p,q) represents the noise-free pixel value at position (p,q) in the original image. a and b are noise parameters. Weak texture sample I is fitted using the least squares method. wk The pixel value-noise variance estimate pair Get the estimated value of the original image noise parameter and After determining the noise estimation value in the image, the size of the noise component in the image can be determined, so as to be used for the subsequent removal of signal-dependent noise of the CMOS image sensor.

Claims

1. A signal-dependent noise parameter estimation method based on improved density peak clustering, characterized by: The specific steps include: Step 1: Use a sliding window to extract several samples of the same size from the original image containing noise; The mean, entropy and gradient of each sample are calculated as the feature data of the sample; the feature data of all samples are normalized to form the feature data set X; Step 2: Calculate the following parameters for the feature data set X: ①Feature data x i The Euclidean distance d between other feature data ij : d ij =||x i -x j || (1) ②Feature data x i The local density ρ i : ③Feature data x i The distance δ to the nearest higher density feature i : ④Characteristic data x i The distance θ to the nearest feature with lower density i : Among them, x i 、x j ∈X,d c is the cutoff distance, and the size of the cutoff distance is the value of 1% to 2% after the distances between all feature data are arranged in ascending order; Step 3: Divide the feature data x according to the calculation results of step 2 i Relative density comparison range η i And calculate its relative density ρ i ′: or i =max(θ i ,d i ) (5) Among them, N i is the relative density comparison range η i The number of feature data in ; Represents feature data x j The weight of Step 4: Use the feature data x obtained in step 3 i The relative density ρ i 'Replace the local density ρ i , recalculate the characteristic data x according to formula (3) i The distance δ to the nearest higher density feature i ; Then calculate the characteristic data x i The clustering value γ i : c i =ρ i '·d i (7) The clustering value γ is calculated i Sort in descending order, select the two feature data with the highest clustering value as the clustering center for clustering, and then output the cluster label corresponding to each feature data to complete the clustering process; Step 5: Calculate the estimated pixel level of the weak texture sample based on the clustering results of step 4 and the noise variance Where N is the side length of the sample, wk represents the index of the weak texture sample, B_weak_n is the number of weak texture samples, I wk (m,n) represents the weak texture sample I wk The pixel value of the mth row and nth column; is the minimum variance direction vector of the weak texture sample, and the covariance matrix C P The smallest eigenvalue of is related to: Step 6: Configure the noise in the original image as a Poisson-Gaussian noise model, then the overall noise method σ at the (p, q) position in the original image 2 (p,q) is: σ 2 (p,q)=ax(p,q)+b (11) Where x(p,q) represents the noise-free pixel value at position (p,q) in the original image; a and b are noise parameters; the weak texture sample I is fitted using the least squares method. wk The pixel value-noise variance estimate pair Get the estimated value of the original image noise parameter and 2. The signal-dependent noise parameter estimation method based on improved density peak clustering as claimed in claim 1, characterized in that: Step 1: Use a sliding window of size 16*16, slide one pixel at a time in order from top to bottom and from left to right, and extract several samples of size 16*16 from the original image containing noise.

3. The signal-dependent noise parameter estimation method based on improved density peak clustering according to claim 1 or 2, characterized in that: For each sample, calculate its mean, entropy and gradient grad as the feature data of the sample: Among them, gray represents the size of the gray value, p(gray) represents the probability that the gray value is gray; W and H are the width and height of the sample respectively, p w,h Represents the pixel value at position (w,h) in the sample.

Images