A 3D Otsu threshold segmentation method based on iteration and dimensionality decomposition

Abstract

The present invention proposes a 3D Otsu threshold segmentation method based on iteration and dimensionality decomposition, which includes: first, using local Laplacian filtering to perform denoising preprocessing on the source image to obtain a smooth layer of the source image; second, designing a 1D Otsu threshold segmentation method based on iteration, iteratively searching the sub-regions in the image, and realizing image threshold segmentation; third, decomposing the three components (pixel gray value, neighborhood mean, neighborhood median) of the 3D Otsu histogram, using the 1D Otsu method based on iteration to search for the optimal thresholds of the pixel gray value, the neighborhood mean, and the neighborhood median respectively, and taking the mean of the three as the final optimal threshold to realize the 3D Otsu threshold segmentation of the image. The method of the present invention can effectively improve the operating efficiency of 3D Otsu through dimensionality decomposition. At the same time, the optimal threshold of 3D Otsu is obtained by an iterative search method to improve the accuracy of image segmentation. Experimental results show that the 3D Otsu threshold segmentation method based on iteration and dimensionality decomposition can effectively segment natural images.

Description

Technical Field

[0001] The invention belongs to the field of image segmentation, and specifically, designs a 3D Otsu threshold segmentation method based on iteration and dimensionality decomposition to improve the efficiency and accuracy of image segmentation. Background Art

[0002] The purpose of image segmentation is to classify each pixel in the image so as to extract the region of interest. As an important step in image processing, image segmentation has attracted widespread attention from scholars. There are many segmentation methods proposed nowadays. As the basis of image processing, the segmentation quality affects the effectiveness of subsequent image processing tasks. The development of image segmentation technology is a process of continuous evolution from manual segmentation to semi-automatic and fully automatic segmentation.

[0003] Researchers have conducted extensive research on image segmentation methods, such as threshold segmentation methods, clustering methods, edge detection-based methods, region growing methods, graph cut methods, and deep learning-based methods. Threshold segmentation is a region-based image segmentation technology. The principle is to divide image pixels into several categories by setting thresholds. The threshold segmentation method is simple to implement, has low computational complexity, and relatively stable performance. It is the most basic and widely used segmentation method for image segmentation. The threshold segmentation method is divided into single threshold and multi-threshold: a single threshold divides the image into foreground and background. When there are multiple targets to be extracted, a single threshold can no longer meet the needs of image segmentation. The multi-threshold method is suitable for situations where there are multiple targets to be extracted and these targets are distributed in different grayscale ranges: for example, some medical images are often composed of black background, white matter area, and gray matter area. The use of multi-threshold segmentation can accurately distinguish the various components in the image, making it easier to extract specific targets, such as lesion areas.

[0004] The threshold segmentation method only considers the grayscale of pixels in the image, and ignores other information of pixels in the image, such as spatial information; in order to make more effective use of the spatial information contained in the image, a three-dimensional threshold segmentation method is proposed; the three-dimensional threshold segmentation method comprehensively utilizes the relevant information between the pixels in the image and their neighborhood space, has good anti-noise performance, and the segmentation effect is more ideal. Correspondingly, the operating efficiency of the three-dimensional threshold segmentation algorithm is low. Therefore, how to propose an efficient and high-precision algorithm has become a major research hotspot. Summary of the invention

[0005] The purpose of image segmentation is to classify each pixel in the image so as to extract the region of interest. As an important step in image processing, image segmentation has attracted widespread attention from scholars. There are many segmentation methods proposed nowadays. As the basis of image processing, the segmentation quality affects the effectiveness of subsequent image processing tasks. The development of image segmentation technology is a process of continuous evolution from manual segmentation to semi-automatic and fully automatic segmentation.

[0006] Researchers have conducted extensive research on image segmentation methods, such as threshold segmentation methods, clustering methods, edge detection-based methods, region growing methods, graph cut methods, and deep learning-based methods. Threshold segmentation is a region-based image segmentation technology. The principle is to divide image pixels into several categories by setting thresholds. The threshold segmentation method is simple to implement, has low computational complexity, and relatively stable performance. It is the most basic and widely used segmentation method for image segmentation. The threshold segmentation method is divided into single threshold and multi-threshold: a single threshold divides the image into foreground and background. When there are multiple targets to be extracted, a single threshold can no longer meet the needs of image segmentation. The multi-threshold method is suitable for situations where there are multiple targets to be extracted and these targets are distributed in different grayscale ranges: for example, some medical images are often composed of black background, white matter area, and gray matter area. The use of multi-threshold segmentation can accurately distinguish the various components in the image, making it easier to extract specific targets, such as lesion areas.

[0007] The threshold segmentation method only considers the grayscale of pixels in the image, and ignores other information of pixels in the image, such as spatial information; in order to make more effective use of the spatial information contained in the image, a three-dimensional threshold segmentation method is proposed; the three-dimensional threshold segmentation method comprehensively utilizes the relevant information between the pixels in the image and their neighborhood space, has good anti-noise performance, and the segmentation effect is more ideal. Correspondingly, the operating efficiency of the three-dimensional threshold segmentation algorithm is low. Therefore, how to propose an efficient and high-precision algorithm has become a major research hotspot.

[0008] The purpose of the present invention is to overcome the shortcomings of the prior art, improve the operating efficiency and accuracy of image segmentation, and propose a 3D Otsu threshold segmentation method based on iteration and dimensionality decomposition, which has better segmentation effect than the traditional threshold segmentation method.

[0009] In order to achieve the above object, the technical solution adopted by the present invention is:

[0010] A 3D Otsu threshold segmentation method based on iteration and dimensionality decomposition, the key steps of which include:

[0011] Step 1: Use local Laplacian filtering to perform denoising preprocessing on the source image to obtain a smoothing layer of the source image;

[0012] Step 2: Design an iterative 1D Otsu threshold segmentation method to iteratively search for sub-regions in the image to achieve image threshold segmentation;

[0013] Step three, decompose the three components of the 3D Otsu histogram (pixel gray value, neighborhood mean, and neighborhood median), use the iterative 1D Otsu method to search for the optimal thresholds of the pixel gray value, neighborhood mean, and neighborhood median respectively, and take the mean of the three as the final optimal threshold to achieve 3D Otsu threshold segmentation of the image.

[0014] The local Laplacian filtering method described in step 1 has the following model:

[0015] Given a source image I , whose size is , the single-scale local Laplacian filter is defined as:

[0016] ;

[0017] Among them, q and p represent the spatial position of the pixel, and represent the source image respectively. I The grayscale value of the pixel at position q and p in , represents the neighborhood of pixel q, , represents the normalized Gaussian kernel, The variance is The Gaussian kernel of , f represents a continuous mapping function, which is usually set to have a variance of The Gaussian kernel of , and f They represent the spatial weight and range weight when building pyramids respectively.

[0018] In step 2, an iterative 1D Otsu threshold segmentation method is designed to iteratively search for sub-regions in the image to achieve image threshold segmentation. The segmentation method is:

[0019] In the first iteration, the segmented image is treated I Perform 1D Otsu threshold segmentation according to I Grayscale histogram of t 1 and the class mean μ 1,0 , μ 1,1 , if the gray value of the pixel , the pixel is divided into C Class 0 (background area); if the gray value of the pixel , the pixel is divided into CCategory 1 (target area); grayscale value in the interval [ μ 1,0 , μ 1,1 ] will be divided in subsequent iterations.

[0020] In the second iteration, in the interval [ μ 1,0 , μ 1,1 ] to perform 1D Otsu threshold segmentation and obtain the threshold t 2 and class mean μ 2,0 , μ 2,1 , if the gray value of the pixel , the pixel is divided into C Class 0 (background area); if the gray value of the pixel , the pixel is divided into C Category 1 (target area); grayscale value in the interval [ μ 2,0 , μ 2,1 ] will continue to be divided in the next iteration.

[0021] Repeat the above iterative process, gradually update the background area and the target area, if the difference between the threshold obtained in the current iteration and the threshold obtained in the previous iteration is less than ,Right now , then stop the iteration, the optimal threshold is .

[0022] In step 3, the three components of the 3D Otsu histogram (pixel gray value, neighborhood mean, and neighborhood median) are decomposed, and the optimal thresholds of the pixel gray value, neighborhood mean, and neighborhood median are searched respectively using the iterative 1D Otsu method, and the mean of the three is used as the final optimal threshold to achieve 3D Otsu threshold segmentation of the image. The segmentation method is:

[0023] For a source image I , the total number of pixels is expressed as ,in, n x Indicates the gray value is x The number of pixels, n y The neighborhood mean is y The number of pixels, n z The median value of the neighborhood is z The number of pixels, n x , ny , n z The probability of , , ;

[0024] Gray value f , neighborhood mean g and the neighborhood median h The between-class variance of the three components is defined as follows:

[0025] in, and Respectively represent q The class probabilities and class means of the classes, Represents the total mean of all classes and is defined as follows:

[0026]

[0027] Maximize the following objective function , and , respectively obtain the optimal thresholds of the three components,

[0028]

[0029] Finally, as the final optimal threshold.

[0030] The main advantages of the method proposed in the present invention are: (1) the local Laplacian filter is used to obtain the smooth layer of the source image, which can effectively reduce the influence of noise on the segmentation result; (2) a 1D Otsu threshold segmentation method based on iteration is designed to obtain more refined segmentation results and improve the accuracy of image segmentation; (3) a 3D Otsu threshold segmentation method based on dimensionality decomposition is designed, and the optimal thresholds of the three components (pixel gray value, neighborhood mean, and neighborhood median) are searched separately by 1D Otsu to achieve fast 3D Otsu threshold segmentation and improve the operating efficiency of image segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the 3D Otsu threshold segmentation method based on iteration and dimensionality decomposition of the present invention.

[0032] Figure 2 This is an example of image segmentation results obtained by the 3D Otsu threshold segmentation method based on iteration and dimensionality decomposition and other segmentation algorithms of the present invention. DETAILED DESCRIPTION

[0033] The present invention is described in detail below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be pointed out that without departing from the core idea of ​​the present invention, those skilled in the art can make some improvements to the present invention, which all belong to the protection scope of the present invention.

[0034] like Figure 1 As shown, the present invention provides a 3D Otsu threshold segmentation method based on iteration and dimensionality decomposition, comprising the following steps:

[0035] Step 1: Use local Laplacian filtering to perform denoising preprocessing on the source image to obtain the smoothing layer of the source image. The method model is:

[0036] Given a source image I , whose size is , the single-scale local Laplacian filter is defined as:

[0037] Among them, q and p represent the spatial position of the pixel, and Represents the source images I The grayscale value of the pixel at position q and p in , represents the neighborhood of pixel q, represents the normalized Gaussian kernel, The variance is The Gaussian kernel of , f represents a continuous mapping function, which is usually set to have a variance of The Gaussian kernel of , and f They represent the spatial weight and range weight when building pyramids respectively.

[0038] Step 2: Design an iterative 1D Otsu threshold segmentation method to iteratively search for sub-regions in the image to achieve image threshold segmentation. The segmentation method is:

[0039] In the first iteration, the segmented image is treated I Perform 1D Otsu threshold segmentation according to I Grayscale histogram of t 1 and the class mean μ 1,0 , μ 1,1 , if the gray value of the pixel , the pixel is divided into C Class 0 (background area); if the gray value of the pixel , the pixel is divided into C Category 1 (target area); grayscale value in the interval [ μ 1,0 , μ 1,1 ] will be divided in subsequent iterations.

[0040] In the second iteration, in the interval [ μ 1,0 , μ 1,1 ] to perform 1D Otsu threshold segmentation and obtain the threshold t 2 and class mean μ 2,0 , μ 2,1 , if the gray value of the pixel , the pixel is divided into C Class 0 (background area); if the gray value of the pixel , the pixel is divided into C Category 1 (target area); grayscale value in the interval [ μ 2,0 , μ 2,1 ] will continue to be divided in the next iteration.

[0041] Repeat the above iterative process, gradually update the background area and the target area, if the difference between the threshold obtained in the current iteration and the threshold obtained in the previous iteration is less than ,Right now , then stop the iteration, the optimal threshold is .

[0042] In step 3, the three components of the 3D Otsu histogram (pixel gray value, neighborhood mean, and neighborhood median) are decomposed, and the optimal thresholds of the pixel gray value, neighborhood mean, and neighborhood median are searched respectively using the iterative 1D Otsu method, and the mean of the three is used as the final optimal threshold to achieve 3D Otsu threshold segmentation of the image. The segmentation method is:

[0043] For a source image I , the total number of pixels is expressed as ,in, n x Indicates the gray value is x The number of pixels, n y The neighborhood mean is y The number of pixels, n z The median value of the neighborhood is z The number of pixels, nx , n y , n z The probability of , , .

[0044] Gray value f , neighborhood mean g and the neighborhood median h The between-class variance of the three components is defined as follows:

[0045] ;

[0046] in, , Respectively represent q The class probabilities and class means of the classes, Represents the total mean of all classes and is defined as follows:

[0047] ;

[0048] Maximize the following objective function , and , respectively obtain the optimal thresholds of the three components,

[0049] ;

[0050] Finally, as the final optimal threshold.

[0051] In this embodiment, Figure 2 is an example of segmentation results obtained by the 3D Otsu threshold segmentation method based on iteration and dimensionality decomposition and other threshold segmentation algorithms of the present invention, Figure 2 Includes two sets of segmentation results; among them, Figure 2 (a1) and Figure 2 (b1) are two source images, Figure 2 (a2) and Figure 2 (b2) are the corresponding standard segmentation results (Ground-truth); it can be seen from the figure that the method provided by the present invention can obtain satisfactory segmentation results and is superior to other segmentation algorithms.

[0052] Table 1 shows the Jaccard measurement evaluation results of different segmentation algorithms, where the best evaluation results are marked in bold. It can be seen from Table 1 that the segmentation results obtained by the 3D Otsu threshold segmentation method based on iteration and dimensionality decomposition provided by the present invention are better than those of other traditional threshold segmentation algorithms.

[0053] Test image The present invention Otsu-IR Global Otsu Direct block Otsu <![CDATA[(a1)]]> 0.9681 0.9656 0.9524 0.7605 <![CDATA[(b1)]]> 0.9847 0.9612 0.9790 0.8363

Claims

1. A 3D Otsu threshold segmentation method based on iteration and dimensionality decomposition, characterized in that: This is accomplished by the following steps: Step 1: Use local Laplacian filtering to perform denoising preprocessing on the source image to obtain a smoothing layer of the source image; Step 2: Design an iterative 1D Otsu threshold segmentation method to iteratively search for sub-regions in the image to achieve image threshold segmentation; In the first iteration, 1D Otsu threshold segmentation is performed on the image I to be segmented. According to the grayscale histogram of I, the initial threshold t1 and the class mean μ are obtained. 1,0 , μ 1,1; For the background area, if the gray value of the pixel p≤μ 1,0 , the pixel is classified into C0 class; for the target area, if the gray value of the pixel p ≥ μ 1,1 , the pixel is classified into C1 class; the gray value is in the interval [μ 1,0 ,μ 1,1 ] will be divided in subsequent iterations; In the second iteration, in the interval [μ 1,0 ,μ 1,1 ] to perform 1D Otsu threshold segmentation and obtain the threshold t2 and class mean μ 2,0 , μ 2,1; For the background area, if the gray value of the pixel p≤μ 2,0 , the pixel is classified into C0 class; for the target area, if the gray value of the pixel p ≥ μ 2,1 , the pixel is classified into C1 class; the gray value is in the interval [μ 2,0 ,μ 2,1 ] will continue to be divided in the next iteration; Repeat the above iterative process and gradually update the background area and the target area; if the difference between the threshold obtained in the current iteration and the threshold obtained in the previous iteration is less than θ (θ>0), that is, |t s -t s-1 |<θ, then stop the iteration, the optimal threshold is T=t s; Step three, decompose the three components of the 3D Otsu histogram: pixel gray value, neighborhood mean, and neighborhood median, and use the iterative 1D Otsu method to search for the optimal thresholds of pixel gray value, neighborhood mean, and neighborhood median respectively, and take the mean of the three as the final optimal threshold to achieve 3D Otsu threshold segmentation of the image.

2. The 3D Otsu threshold segmentation method based on iteration and dimensionality decomposition according to claim 1, characterized in that: The local Laplacian filtering method described in step 1 is described as follows: Given a source image I of size M×N, the single-scale local Laplacian filter is defined as: Where q and p represent the spatial position of the pixel; I p and I q Respectively represent the grayscale values ​​of pixels at positions q and p in the source image I; Ω p represents the neighborhood of pixel q; represents the normalized Gaussian kernel, The variance is The Gaussian kernel of f represents a continuous mapping function, which is usually set to have a variance σ 2 Gaussian kernel, that is, f(x) = G σ (x) = exp(-x 2 / 2σ 2 ); and f represent the spatial weight and range weight when building the pyramid, respectively.

3. The 3D Otsu threshold segmentation method based on iteration and dimensionality decomposition according to claim 1, characterized in that: In step three, the three components of the 3D Otsu histogram are decomposed: pixel gray value, neighborhood mean, and neighborhood median. The optimal thresholds of pixel gray value, neighborhood mean, and neighborhood median are searched respectively by the iterative 1D Otsu method, and the mean of the three is used as the final optimal threshold to realize the 3D Otsu threshold segmentation of the image; the segmentation method is: For a source image I, the total number of pixels is expressed as Among them, n x Indicates the number of pixels with gray value x, n y Indicates the number of pixels whose neighborhood mean is y, n z The number of pixels with value z in the neighborhood; n x , n y , n z The probability of The inter-class variance of the three components of gray value f, neighborhood mean g and neighborhood median h is defined as follows: Among them, ω q and μ q Respectively represent the class probability and class mean of the qth (q=0,1) class; μ T Represents the total mean of all classes and is defined as follows: Maximize the following objective function and Get the optimal thresholds of the three components respectively, Finally, as the final optimal threshold.

Images