Image segmentation method and system based on feature driven heuristic four-color label

An image segmentation, heuristic technique, applied in the field of computer vision

Active Publication Date: 2018-03-23
SOUTH CENTRAL UNIVERSITY FOR NATIONALITIES
6 Cites 2 Cited by

AI-Extracted Technical Summary

Problems solved by technology

Another obvious weakness of the random coloring strategy is: assigning the sam...
View more

Abstract

The invention discloses an image segmentation method and system based on a feature driven heuristic four-color label, and relates to the field of machine vision. The method comprises the following steps that a meanshift method is used, and initialized segmentation is carried out on an input image; global grouping is carried out on the image after initialized segmentation, distribution of an initial area of a feature space is analyzed, and a similar matrix in an area set of initialized segmentation serves as AP clustering input; unnecessary adjacency is cracked via an adjacent relation crackingalgorithm, so that uniform adjacent areas can be marked in the same color; a heuristic four-color label algorithm is used to establish an internal coloring relation adaptively; and an MMPC model is combined with a GAC model to establish an MMPC-GAC model, MMPC-GAC modeling and MLG optimization are carried out iteratively till convergence is reached, and a final four-color segmented image is obtained. According to the invention, the uniform adjacent areas can be marked in the same color, and a uniform appearance area becomes globally consistent.

Application Domain

Image analysis

Technology Topic

Image segmentationFeature based +6

Image

  • Image segmentation method and system based on feature driven heuristic four-color label
  • Image segmentation method and system based on feature driven heuristic four-color label

Examples

  • Experimental program(1)

Example Embodiment

[0042] The present invention will be further described in detail below in conjunction with the drawings and specific embodiments.
[0043] See figure 1 As shown, the embodiment of the present invention provides an image segmentation method based on feature-driven heuristic four-color labels, including the following steps:
[0044] S1. Use the mean shift method based on clustering to initialize the input image, where the mean shift parameter setting has been fixed;
[0045] The embodiment of the present invention uses images of BSDS300 and its extended version BSDS500 whose size is fixed at 481×321 (321×481), mean shift is used as an over-segmentation method for initialization, and its parameters are set;
[0046] Globally group the images after the initial segmentation, analyze the distribution of the initial regions of the feature space, and use the similarity matrix on the initial segmented region set as the AP (Affinity Prorogation) clustering input; use the neighbor relationship cracking algorithm to crack Unnecessary adjacency, so that evenly adjacent areas can be marked with the same color;
[0047] S2, using heuristic four-color labeling algorithm to adaptively establish internal coloring relationships;
[0048] S3. Combine the MMPC (Multiphase Multiple Piecewise Constant) model and the GAC (Geodesic Active Contour) model to establish an MMPC-GAC model, describe the target and background of the unevenness, and iterate MMPC-GAC modeling and multi-layer graph MLG optimization until convergence is reached, and the final four-color segmentation image is obtained.
[0049] Step S1 specifically includes the following steps:
[0050] See figure 2 As shown, the input image is preliminarily segmented using the mean shift algorithm to obtain a preliminary segmented region set R. R is a set divided into several regions, R={r i Or r j }, i and j are positive integers, where r i , R j Is the area number, A ij Is the initial adjacency matrix, when the region r i And r j When adjacent, A ij =1, otherwise 0; S ij Is the similarity matrix on the preliminary segmentation area set R; the adjacent relationship cracking algorithm is used, and the broken adjacent matrix output by the algorithm is A ij ’, initialization makes A ij ’=A ij , Use AP clustering to crack the similarity matrix S ij Unnecessary adjacency, so that evenly adjacent areas can be marked with the same color; if r i And r j Belong to the same cluster, reset A ij '=0, to update the adjacent matrix.
[0051] Step S2 specifically includes the following steps:
[0052] See figure 2 As shown, the heuristic four-color labeling algorithm is used to initialize the adjacent matrix A ij , Color area collection C l , Color area set C l The initial stage is an empty set, initialize the colored label Available color indicator a l i =1, color order o l i =l, where l is one of four different color labels, and i is the corresponding area;
[0053] From left to right, from top to bottom to area r i Number, and then color according to the number sequence; And push the first area into color area C 1 After that, start the coloring cycle from the second area;
[0054] In the area r i Before assigning colors, first evaluate the area r i And all colored areas set C l Characteristic distance between d l i , According to the characteristic distance d l i Sort the candidate colors in ascending order, the order of color label l will be stored in o l i Medium, to provide consistent colors for uniform areas;
[0055] According to o l i Try all colors, once the current color label l meets "available": a l i = 1 and adjacent constraints: A i, j When ≠1, the area Will be assigned to the color label l; then, r i Push into colored area C l , Now set a l i =0, the area r i The color label l marked as unavailable, move to the next area; if there is no satisfactory color, recolor the previous area.
[0056] In step S2, in the coloring process, a l i =1 means the color label l can be used in the area r i , Once the color label l has been tested or used, a l i Will be set to 0, indicating that the color label l is not available.
[0057] Step S3 specifically includes the following steps:
[0058] Perform an iterative loop on the input source image I to achieve MMPC-GAC modeling and MLG optimization, and treat the area and pixels with the same color or the same color label as a phase P l , Label function p is the pixel of the source image I; each phase is grouped into sub-phases with K-means, and the multi-segment constant function is calculated; the MLG method is used to iteratively solve the multi-stage optimization problem, and the pixel label is obtained. When it is changed again or the number of iterations reaches the upper limit, the iteration ends, and the final four-color segmented image is obtained.
[0059] The embodiment of the present invention also provides an image segmentation system based on feature-driven heuristic four-color labels, the system including an initialization segmentation unit, a coloring unit, and an iterative optimization unit;
[0060] The initial segmentation unit is used to: use the mean shift method based on clustering to initialize the input image; group the images after the initial segmentation, analyze the distribution of the initial area of ​​the feature space, and group the initially segmented area The similarity matrix of is used as the input of affinity AP clustering; the neighbor relationship cracking algorithm is used to crack unnecessary adjacencies, so that evenly adjacent areas can be marked with the same color;
[0061] The coloring unit is used to: adopt a heuristic four-color label algorithm to adaptively establish the internal coloring relationship;
[0062] The iterative optimization unit is used to combine the multiphase multi-segment constant MMPC model and the geodetic active contour GAC model, establish the MMPC-GAC model, describe the target and background of the inhomogeneity, and iteratively perform the MMPC-GAC modeling and multi-layer graph MLG optimizes until convergence is reached, and the final four-color segmented image is obtained.
[0063] The initial segmentation unit uses the mean shift algorithm to preliminarily segment the input image to obtain a preliminary segmentation region set R, R is a set divided into several regions, R={r i Or r j }, i and j are positive integers, where r i , R j Is the area number, A ij Is the initial adjacency matrix, when the region r i And r j When adjacent, A ij =1, otherwise 0; S ij Is the similarity matrix on the preliminary segmentation area set R; the adjacent relationship cracking algorithm is used, and the broken adjacent matrix output by the algorithm is A ij ’, initialization makes A ij ’=A ij , Use AP clustering to crack the similarity matrix S ij Unnecessary adjacency, so that evenly adjacent areas can be marked with the same color.
[0064] If r i And r j Belong to the same cluster, initialize the split unit and reset A ij '=0, update the adjacent matrix.
[0065] The coloring unit uses heuristic four-color labeling algorithm to initialize the adjacent matrix A ij , Color area collection C l , Color area set C l The initial stage is an empty set, initialize the colored label Available color indicator a l i =1, color order o l i =l, where l is one of four different color labels, and i is the corresponding area;
[0066] From left to right, from top to bottom to area r i Number, and then color according to the number sequence; And push the first area into color area C 1 After that, start the coloring cycle from the second area;
[0067] In the area r i Before assigning colors, first evaluate the area r i And all colored areas set C l Characteristic distance between d l i , According to the characteristic distance d l i Sort the candidate colors in ascending order, the order of color label l will be stored in o l i Medium, to provide consistent colors for uniform areas;
[0068] According to o l i Try all colors, once the current color label l meets "available": a l i = 1 and adjacent constraints: A i, j When ≠1, the area Will be assigned the color label l; then, r i Push into colored area C l , Now set a l i =0, the area r i The color label l marked as unavailable, move to the next area; if there is no satisfactory color, recolor the previous area.
[0069] The iterative optimization unit performs an iterative loop on the input source image I, realizes MMPC-GAC modeling and MLG optimization, and treats areas and pixels with the same color or the same color label as a phase P l , Label function p is the pixel of the source image I; each phase is grouped into sub-phases with K-means, and the multi-segment constant function is calculated; the MLG method is used to iteratively solve the multi-stage optimization problem, and the pixel label is obtained. When it is changed again or the number of iterations reaches the upper limit, the iteration ends, and the final four-color segmented image is obtained.
[0070] The embodiment of the present invention is faced with complex natural landscape pictures and can display more reasonable color maps. In images with clutter and complex structure, the iterative algorithm will produce better segmentation and faster convergence. In repeated experiments, compared with the random four-color labeling method, the heuristic four-color label can obtain a better number of segmented images, and is more suitable for complex and difficult situations. Another feature is that it is usually in a low number of iterations ( For example, 2-3 iterations) produce better segmentation.
[0071] In the quantitative comparison with the most advanced method, the BSDS300 database is used for comparison, and the AP clustering preference value in the embodiment of the present invention is fixed to the median value of the similarity matrix. The embodiment of the present invention obtains a competitive advantage under the measurement of PRI (Probabilistic RandIndex) and GCE (Global Consistency Error), even in VoI (Variation of Information) and BDE (Boundary Displacement Error). , Boundary displacement error), the performance is better, especially the measurement results of RFCL (Random Four-Color Labeling) on ​​the R Better (Random Better) group and HBetter (Heuristic Better, heuristic) When the measurements of HFCL (Heuristic Four-Color Labeling) in the better) group are combined, it is found that it is almost superior to all comparison methods under all metrics.
[0072] The embodiment of the present invention is a good alternative to the random four-color labeling strategy, especially when the random four-color label (RFCL) does not perform well in complex scenes involving images.
[0073] Those skilled in the art can make various modifications and variations to the embodiments of the present invention. If these modifications and variations fall within the scope of the claims of the present invention and their equivalent technologies, these modifications and variations are also within the protection scope of the present invention. .
[0074] The content not described in detail in the specification is the prior art known to those skilled in the art.

PUM

no PUM

Description & Claims & Application Information

We can also present the details of the Description, Claims and Application information to help users get a comprehensive understanding of the technical details of the patent, such as background art, summary of invention, brief description of drawings, description of embodiments, and other original content. On the other hand, users can also determine the specific scope of protection of the technology through the list of claims; as well as understand the changes in the life cycle of the technology with the presentation of the patent timeline. Login to view more.
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products