CN111598890A - Level set optimization method for underwater image segmentation - Google Patents
Level set optimization method for underwater image segmentation Download PDFInfo
- Publication number
- CN111598890A CN111598890A CN202010410102.6A CN202010410102A CN111598890A CN 111598890 A CN111598890 A CN 111598890A CN 202010410102 A CN202010410102 A CN 202010410102A CN 111598890 A CN111598890 A CN 111598890A
- Authority
- CN
- China
- Prior art keywords
- level set
- function
- term
- model
- underwater image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 53
- 238000005457 optimization Methods 0.000 title claims abstract description 20
- 238000003709 image segmentation Methods 0.000 title claims abstract description 17
- 238000009792 diffusion process Methods 0.000 claims abstract description 10
- 238000013459 approach Methods 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 230000015572 biosynthetic process Effects 0.000 claims description 2
- 230000003750 conditioning effect Effects 0.000 claims description 2
- 238000010276 construction Methods 0.000 claims description 2
- 238000003786 synthesis reaction Methods 0.000 claims description 2
- 230000011218 segmentation Effects 0.000 abstract description 22
- 230000002829 reductive effect Effects 0.000 abstract description 2
- 230000000694 effects Effects 0.000 description 6
- 230000001965 increasing effect Effects 0.000 description 4
- 230000010355 oscillation Effects 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 230000002708 enhancing effect Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 108091034117 Oligonucleotide Proteins 0.000 description 1
- 230000002411 adverse Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000012733 comparative method Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000009499 grossing Methods 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000036961 partial effect Effects 0.000 description 1
- 230000000452 restraining effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a level set optimization method for underwater image segmentation, which mainly carries out optimization design on an internal energy function of a level set, namely a regular term, and improves the diffusivity of a regular term diffusion function so as to accelerate the convergence speed and stability of a model; the level set function is normalized through the proposed regular term, so that the reinitialization problem of the model is solved while the character of the symbolic function is maintained, and the segmentation precision lost due to reinitialization is reduced. Compared with the traditional level set method, the improved method has the advantages of higher convergence rate, higher segmentation precision and better robustness for underwater image segmentation.
Description
Technical Field
The invention belongs to the technology in the field of image processing, and relates to a level set optimization method for underwater image segmentation.
Background
The reinitialization problem of the level set adds a great computational cost to the model, the model generates oscillation in the process and the segmentation precision is easily lost, and nowadays, many models are proposed for processing the reinitialization problem of the level set. However, due to the influence of an underwater complex optical environment, the difficulty of underwater image segmentation is greater than that of land image segmentation, and how to overcome the interference of noise in an underwater image on a model becomes a big problem. The conventional regular term model theory is simple in hypothesis and difficult to be used for underwater image segmentation, and noise in an underwater complex scene degrades the effect of a regular term, so that the convergence speed of a level set model is slowed down and oscillation is generated.
Oher and Sethian proposed the concept of level set implicit in 1988; malladi and Caselles et al apply this concept in image segmentation in turn; chopp, based on caseles, proposes an energy function and incorporates the level set model in the form of a constraint term. However, the level set function needs to be kept as a symbol distance function in the evolution process, so the level set function needs to be periodically reinitialized to maintain the performance of the model. But the reinitialization of the level set takes a lot of time, increasing the computation cost of the level set function. Based on the M-S model, Chan and Vese et al introduced a homogeneous hypothesis theory and proposed a famous global segmentation model C-V model. The objective function of the model is non-convex and the model still needs to be reinitialized to preserve the regularity of the level set function evolution. Li et al introduced a regularization function based on the model proposed by Caselles, which is essentially based on the level set function of the edge, and after the regularization term is fused, the level set function is kept as a symbolic distance function in the evolution process, so that the model is free from reinitialization and the convergence speed is increased, and the model is represented by a DRLSE1 model. However, the model is unstable in value under special conditions and is easy to have negative influence on the model. In 2010, Li et al proposed an optimized distance regularization level set model to overcome adverse effects on the basis of previous work, enhancing the segmentation performance of the model, which is denoted by DRLSE 2. Zhang et al later proposed a gaussian regularization method, which was introduced into a level set model to achieve local segmentation, and the evolution process of the model was equivalent to the gaussian smoothing process, thereby avoiding the reinitialization process in the level set evolution. It is worth noting that the model can segment the target object quickly, but there is no further investigation on model oscillation. Then Zhang et al further proposed a new diffusion (RD) method for implicit active contours without re-initialization during the evolution process. The method has good performance on images which are divided into weak edges and easy to generate boundary leakage, but the convergence rate of the method is lower than that of a Gaussian regular level set model.
Disclosure of Invention
The purpose of the invention is as follows: the invention aims to provide a level set optimization method for underwater image segmentation, which improves the stability and convergence speed of model evolution by improving the diffusivity of a diffusion function in a regular term, solves the problems of reinitialization and slow convergence of a level set and further optimizes the segmentation effect of an underwater image.
The technical scheme is as follows: in order to achieve the above object, the present invention provides a level set optimization method for underwater image segmentation, comprising the following steps:
(1) the regular term is optimally designed, and the mode of the integrand function in the regular term is the gradientIs a variable, inAndin two intervals, withTo constrainApproach to 1 and lie inIn the region of (1), utilizeAndmultiply and accelerateA speed approaching 1;
(2) construction of level set energy functional EZ=EW+ENWherein the regularization term ENInternal energy terms, data items E, forming energy functionalWAn external energy term constituting an energy functional;
(3) initial level set function of phi0=cos(r1x)cos(r2y) in which r1Is the product of the input underwater image height and pi, r2The product of the width of the input underwater image and pi, wherein (x, y) is the coordinate of a pixel point in the underwater image;
(4) according toUpdating a level set function, and increasing the diffusivity of the function by using the proposed regular term in the evolution process to accelerate the convergence speed of the model, wherein t is time, delta t represents a time step, and n represents iteration times;
(5) and (4) judging whether the level set function phi reaches a convergence condition, and if the level set function phi does not reach the convergence condition, skipping to the step (4) to continue evolving the level set.
The optimization regularization term in the step (1) aims to solve the problem of reinitialization and improve convergence rate, and the function of the optimization regularization term is defined as:
where Ω represents the entire image domain.
solving an evolution equation by using an Euler-Lagrange equation for the optimization regularization term:
the diffusivity function in the optimization regularization term is:
the energy functional synthesis external term E in the step (2)WAnd an internal item ENThe latter overall function EZComprises the following steps:
wherein,as data items, ωiIs a conditioning parameter, S is a saliency value extracted from an underwater image, saiIs the average of the saliency inside and outside the contour, i ═ 1,2 is used as a subscript to distinguish between the inside and outside regions of the contour, g represents the edge weight, and H (Φ) is the hervesseld function.
The level set function is solved by a gradient descent method to obtain an evolution equation:
and (4) updating and iterating according to the level set evolution equation.
Has the advantages that: the method improves the regular term aiming at the problems of reinitialization and slow convergence, improves the diffusivity of the diffusion function of the regular term, accelerates the convergence speed of the model, and reduces the convergence times required by the curve convergence to the target boundary; and the level set function is regularized through the proposed regular term, the reinitialization problem of the model is solved, and the segmentation precision lost due to reinitialization is further reduced. Experiments show that the level set method provided by the invention can accurately realize underwater image segmentation. Compared with the existing level set method, the method has higher segmentation precision and faster convergence rate for the underwater image.
Drawings
FIG. 1 is a flow chart of an underwater image segmentation method in the present invention.
Fig. 2 is a graph of the underwater original image (a) and the segmentation result (b) when there is no regularization term.
Fig. 3 is a diagram of an underwater original image (a) and a division result (b) when a regularization term is present.
FIG. 4 is a graph comparing the diffusivity equation for the regularization term in the present invention with the diffusivity equation for the regularization term in the DRLSE2 model. The dashed line represents the DRLSE2 diffusivity equation forCurve (c) of (d). The solid line represents the diffusivity equation for the model of the present invention with respect toCurve (c) of (d).
FIG. 5 is a zero level set contour result graph of the present invention method and several comparative methods for underwater image segmentation. Where (a) is the underwater original image, (b) is the result of the DRLSE2 method, (c) is the result of the RD method, and (d) is the result of the method of the present invention, and (1) - (8) lines represent underwater images of different objects.
FIG. 6 is a graph of the binarization results of the method of the present invention and several comparison methods for underwater image segmentation. In which (a) the underwater original image, (b) the artificially labeled true values, (c) the result of the DRLSE2 method, (d) the RD method, and (e) the results of the inventive method, the lines (1) to (8) representing underwater images of different objects.
Detailed Description
In order to clearly highlight the objects and advantages of the present invention, the present invention will be further described below with reference to the accompanying drawings in the present embodiment, and an implementation process of a level set optimization method for underwater image segmentation disclosed in the present embodiment mainly includes the following steps:
(1) the underwater image is inputted as any one of the images in the (a) column in fig. 5. The optimized regular term function designed by the invention and the image characteristics extracted from the underwater image respectively form an internal energy function and an external energy function of the level set energy functional, wherein the external energy function is formed by the region level characteristics and the edge level characteristics. Where region-level features refer to the intensity of the image. Edge level features refer to the gradients of an image.
The expression of the level set model framework is as follows:
EZ=EW+EN
in the formula, EWIs a constraint term representing an external energy term. ENRepresents an internal energy function, i.e. a regularization term.
Internal energy function E in the level set modelNThe expression is as follows:
whereinThe modulus representing the gradient, (x, y) are the coordinates of the pixel points in the underwater image, Ω represents the entire image domain,the functional expression is:
in thatAndon the two intervals, the number of the interval,all are provided withThis term is used to constrainGradient ofApproaching to 1. In addition, the models are not used singlySuch a constraint term is asOn the interval of (1), model utilizationThis term andmultiply so thatThe approach to 1 is faster.
Internal energy function E in the level set modelNBy a regularization termAnd (3) solving to obtain a modeling equation:
according to the regular term function, the diffusivity equation can be obtained as follows:
to obtain:
combining the diffusivity equation with fig. 4, we find the relevant properties of the diffusivity equation as follows:
1) in thatThe diffusivity function value h is greater than 0, the function is expressed as forward diffusion, and the model is a gradientSuppressed to 1.
2) In thatIs less than 0, then the function is expressed as back diffusion and the model will beIncreasing to 1.
3) In thatThe diffusivity function value h is greater than 0, the function is expressed as forward diffusion, and the model will beSuppressed to 0. As described above, the level set function can be increased or decreased by such adaptationAnd minimization of the energy function, which in turn drives the model curve to shrink to the desired target boundary.
As shown in FIG. 4, the diffusivity function of the regularization term of the present invention is compared with that of the DRLSE2 model by h2And h1It is shown, to conclude:
1) in thatIn the interval, the positive and negative of the function values are forward diffusion, and the function value h2Greater than h1Represents the proposed function pairThe restraining force is greater than DRLSE2, which makesThe speed of the model is faster, namely the model has a faster convergence speed when being used for segmenting the underwater image, and meanwhile, the model only needs less iteration times.
2) In thatIn the interval, all belong to backward diffusion, and the function value h2Is still greater than h1Similarly, the proposed function pairIs greater than DRLSE2, such thatIs faster.
3) For the intervalBoth the proposed function and the DRLSE2 model guarantee the symbolic distance function property away from the zero level set
The external energy function E in the level set modelWExpressed as:
in the model, g represents the edge weight and the level set function phi represents the curve profile. OmegaiIs an adjustment parameter, H (phi) denotes the Hervesseld function, S is a saliency value extracted from an underwater image, saiIs the average of the saliency inside and outside the contour, i is used as a subscript to distinguish the inside and outside regions of the contour, and (x, y) is the pixel point coordinates in the underwater image.
The expression of the edge detector in the level set model is as follows:
wherein I is an input image, GσIs a gaussian kernel with a standard deviation of sigma,for gradient operation, a convolution operation is denoted.
The Heaviside function is:
the region features in the level set model are extracted by a significance method:
S(x,y)=|I(x,y)-IGass(x,y)|
wherein I is an input image, IGassRepresenting the image after processing with a gaussian filter.
In the region feature extraction method IGassIs defined as:
IGass(x,y)=κζ*I(x,y)
extracting k in the Gaussian image equationζIs a gaussian kernel with a standard deviation ζ.
The region characteristics in the level set model are extracted by a significance method, and the expression of the significance inside and outside the outline is as follows:
and finally, the energy functional of the level set model is as follows:
(2) initializing a level set function, evolving a level set model, and enhancing the segmentation performance of the model on the underwater image by using the optimized regular term.
The initial level set function is:
φ0=cos(r1x)cos(r2y)
in the initial function r1For inputting the product of the height of the underwater image and pi, r2The product of the width of the underwater image and pi is input, and (x, y) is the coordinate of a pixel point in the underwater image.
The model is solved by a gradient descent method:
and obtaining the following by calculating the partial derivative of the energy function:
wherein div represents the divergence, and solving the energy functional to obtain a level set evolution equation:
the present invention uses a quantitative standard Jaccard Similarity (JS) standard to accurately evaluate the segmentation performance of these methods. JS Standard, theoretically two sets P1And P2The ratio of the intersection of (a) to their union, the formula is:
the closer the JS value is to 1, the more P is represented1The closer to P2The higher the segmentation accuracy of the model. P1Represents the result of the segmentation, and P2Representing the standard value of the manual annotation.
The selected underwater image samples are divided into two types in terms of characteristics, one type is represented by fuzzy image details and less background interference, and the degree of distinction between a target and a background in the images is lower; one category appears to be strong noise, and the objects in such images are often mixed with complex background interference, which presents a great challenge to the segmentation model.
TABLE 1 quantitative comparison table of level set model segmentation results
It can be seen from fig. 2 and fig. 3 that the segmentation result of the present invention is significantly better than the segmentation result of the model without the regularization term, and although the target region is extracted from the result of the model without the regularization term, the model without the regularization term has a large amount of point noise, and the model is not reinitialized by the regularization term of the present invention, so that the accuracy is not lost.
Qualitative results fig. 5 further illustrates that the present invention achieves its theoretical effect, has better convergence, and has better segmentation effect on underwater images. In order to fully illustrate the auxiliary effect of the proposed regularization term on the model, the invention quantifies and compares the results of qualitative experiments. From table 1 it can be seen that the JS value score of the present invention is first, illustrating the superiority of the method of the present invention over the two comparison algorithms. In addition, the calculation time and convergence times of the model evolution are counted. The run-time results are shown in Table 1, where the run-times are the CPU time required for the model to complete convergence, and it can be seen that the time spent by the model of the present invention is also ranked first.
In addition, on the premise of ensuring that each algorithm obtains the optimal result, comparing the iteration times of completing the segmentation of the algorithms, the number of the DRLSE2 models is 320, the number of the RD models is 400, and the iteration times of the method is 40. The RD algorithm, although more accurate than the DRLSE2 method in segmentation, iterates slower on this model. The method is obtained by integrating two indexes of the segmentation precision JS and the iteration times, is superior to a comparison method, and can prove the excellent performance of the method.
Claims (6)
1. A level set optimization method for underwater image segmentation is characterized by comprising the following steps: comprises the following steps:
(1) the regular term is optimally designed, and the mode of the integrand function in the regular term is the gradientIs a variable, inAndin two intervals, withTo constrainApproach to 1 and lie inIn the interval of (1), useAndmultiply and accelerateA speed approaching 1;
(2) construction of level set energy functional EZ=EW+ENWherein the regularization term ENIs an internal energy term of an energy functional, data item EWIs an external energy term of the energy functional;
(3) initial level set function of phi0=cos(r1x)cos(r2y) in which r1Is the product of the height of the input underwater image and pi, r2The product of the width of the input underwater image and pi, wherein (x, y) is the coordinate of a pixel point in the underwater image;
(4) according to the equationUpdating a level set function, improving the function diffusivity by using an optimization regular term in the evolution process, and accelerating the convergence speed of the model, wherein t is time, delta t represents a time step, and n represents iteration times;
(5) and (4) judging whether the level set function phi reaches a convergence condition, and if the level set function phi does not reach the convergence condition, skipping to the step (4) to continue evolving the level set.
5. the method of claim 2, wherein the level set optimization is performed by: the energy functional synthesis external term E in the step (2)WAnd an internal item ENThe latter overall function EZComprises the following steps:
wherein,as data items, ωiIs a conditioning parameter, S is a saliency value extracted from an underwater image, saiIs the average of the saliency inside and outside the contour, i ═ 1,2 is used as a subscript to distinguish between the inside and outside regions of the contour, g represents the edge weight, and H (Φ) is the hervesseld function.
6. The method of claim 5, wherein the level set optimization comprises: after the level set function is solved by a gradient descent method, a level set evolution equation is obtained as follows:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010410102.6A CN111598890B (en) | 2020-05-15 | 2020-05-15 | Level set optimization method for underwater image segmentation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010410102.6A CN111598890B (en) | 2020-05-15 | 2020-05-15 | Level set optimization method for underwater image segmentation |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111598890A true CN111598890A (en) | 2020-08-28 |
CN111598890B CN111598890B (en) | 2022-08-26 |
Family
ID=72183715
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010410102.6A Active CN111598890B (en) | 2020-05-15 | 2020-05-15 | Level set optimization method for underwater image segmentation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111598890B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113093198A (en) * | 2021-03-10 | 2021-07-09 | 河海大学 | Acoustic imaging detection method based on multi-scale Markov random field |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20060158447A1 (en) * | 2005-01-14 | 2006-07-20 | Mcgraw Tim | System and method for fast tensor field segmentation |
CN105976378A (en) * | 2016-05-10 | 2016-09-28 | 西北工业大学 | Graph model based saliency target detection method |
CN106296695A (en) * | 2016-08-12 | 2017-01-04 | 西安理工大学 | Adaptive threshold natural target image based on significance segmentation extraction algorithm |
CN110136146A (en) * | 2019-05-17 | 2019-08-16 | 浙江理工大学 | SAR image Watershed segmentation method based on sinusoidal SPF distribution and Level Set Models |
CN111105430A (en) * | 2019-11-28 | 2020-05-05 | 青岛大学 | Variation level set image segmentation method based on Landmark simplex constraint |
-
2020
- 2020-05-15 CN CN202010410102.6A patent/CN111598890B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20060158447A1 (en) * | 2005-01-14 | 2006-07-20 | Mcgraw Tim | System and method for fast tensor field segmentation |
CN105976378A (en) * | 2016-05-10 | 2016-09-28 | 西北工业大学 | Graph model based saliency target detection method |
CN106296695A (en) * | 2016-08-12 | 2017-01-04 | 西安理工大学 | Adaptive threshold natural target image based on significance segmentation extraction algorithm |
CN110136146A (en) * | 2019-05-17 | 2019-08-16 | 浙江理工大学 | SAR image Watershed segmentation method based on sinusoidal SPF distribution and Level Set Models |
CN111105430A (en) * | 2019-11-28 | 2020-05-05 | 青岛大学 | Variation level set image segmentation method based on Landmark simplex constraint |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113093198A (en) * | 2021-03-10 | 2021-07-09 | 河海大学 | Acoustic imaging detection method based on multi-scale Markov random field |
CN113093198B (en) * | 2021-03-10 | 2023-07-25 | 河海大学 | Acoustic imaging detection method based on multi-scale Markov random field |
Also Published As
Publication number | Publication date |
---|---|
CN111598890B (en) | 2022-08-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2021017361A1 (en) | Template matching algorithm based on edge and gradient feature | |
US20220414891A1 (en) | Method for automatic segmentation of fuzzy boundary image based on active contour and deep learning | |
US9959466B2 (en) | Object tracking apparatus and method and camera | |
CN111681300B (en) | Method for obtaining target area composed of outline sketch lines | |
CN110120057A (en) | Fuzzy region sexuality contours segmentation model based on the global and local fitting energy of weight | |
CN117173201B (en) | Second order differential image segmentation method, system, medium and device | |
CN117593323B (en) | Image segmentation method, system, medium and device based on non-local features | |
CN111598890B (en) | Level set optimization method for underwater image segmentation | |
Yang | Local smoothness enforced cost volume regularization for fast stereo correspondence | |
Huang et al. | Variational level set method for image segmentation with simplex constraint of landmarks | |
Badshah et al. | On two multigrid algorithms for modeling variational multiphase image segmentation | |
CN111105430A (en) | Variation level set image segmentation method based on Landmark simplex constraint | |
Bao et al. | Solar panel segmentation under low contrast condition | |
CN114596320A (en) | Image segmentation method and device based on ALSLCV model | |
Sanou et al. | Semi-automatic tools for nanoscale metrology and annotations for deep learning automation on electron microscopy images | |
Zia et al. | Active Contour Model for Image Segmentation | |
CN112700459A (en) | Level set infrared image segmentation method based on multi-feature information fusion | |
Zhang et al. | Divergence of gradient convolution: deformable segmentation with arbitrary initializations | |
Xue et al. | A novel active contour model based on features for image segmentation | |
CN111209532A (en) | Mixed Gaussian model parameter estimation algorithm based on enhanced online search principle | |
CN111640115B (en) | Non-uniform image segmentation method based on offset field and global smooth prior | |
CN117994277B (en) | Target segmentation method, device, medium and system for non-uniform illumination image | |
Mustaffa et al. | Some numerical methods for solving geodesic active contour model on image segmentation process | |
Wang et al. | Full-image guided method based on confidence coefficient for fast stereo matching | |
El Hassouni et al. | A new weighted normal-based filter for 3D mesh denoising |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |