A kind of multi-modality images dividing method based on Functional Mapping
Technical field
The invention belongs to the technical field of image segmentation in image procossing, particularly based on the multi-modality images dividing method of Functional Mapping.
Background technology
The flourish of Digital image technology has expedited the emergence of a large amount of new industries, as remote sensing satellite image location, medical image analysis, traffic intelligent identification etc., facilitates the ripe day by day of informationized society.Image, as the important bridge in the human perception world, is also closely related with visual field.Such as, the demand of image procossing in all kinds of vision application in the fields such as artificial intelligence, machine vision, physiology, medical science, meteorology, military science constantly increases, and plays more and more crucial effect.And Iamge Segmentation is as image pre-processing method, for solid foundation has been established in image semantic analysis on the middle and senior level, such as, in many application such as image recognition, target localization, rim detection, all image Segmentation Technology improving performance can be used.
Iamge Segmentation, as the term suggests be that given image is carried out region segmentation according to certain rule or target.Such as, the image of a secondary lakeside shooting can be divided into the region of the different semantic classess of multiple representative such as lake surface, people, canoe, house, grove, sky, and people here and canoe can regard target foreground objects as, and remaining can regard background object as.Traditional image Segmentation Technology processes mainly for the clue such as gray scale, color, texture, shape of single image, and typical method has Threshold segmentation, region segmentation, edge segmentation, figure to split, based on energy functional segmentation etc.Such as threshold segmentation method judges its generic according to setting threshold value to gray-scale value; Edge segmentation method has the feature such as step evolution or mutability according to edge gray-scale value and detects; Region segmentation method judges according to image similarity criterion, mainly contains the technology such as watershed divide, regional merge and split, a period of time region growing; Image to be regarded as and is summit with pixel and non-directed graph that neighbor connects with limit by figure segmentation, and each cut zone regards the subgraph in figure as; Utilize continuous curve to represent object edge based on energy functional segmentation, and minimized by energy functional and solve segmentation result, be generally divided into parametric active contour model and two kinds, geometric active contour model.
The not foot point of said method is mainly manifested in the following aspects: the first, directly for the process of image original pixels, adds the time complexity of algorithm, increases computing cost; The second, bottom layer treatment technology such as Threshold segmentation and edge segmentation are difficult to be associated with the semantic feature of image; 3rd, have ignored the complementary information between image, there are some common structure and potential informations between the image especially comprising similar purpose, directly affects the segmentation effect of image object.Therefore, these methods are also not suitable for the Iamge Segmentation task comprising common objective on a large scale, and the practical application such as image recognition, target localization larger to the order of magnitude thus produces certain adverse effect.Consider based on these, for applications such as intelligent transportation identification, medical science impact analysis, large-scale image identifications, a kind ofly can set up associating of image low-level image feature and semantic feature in the urgent need to designing, and multi-facetedly effectively can utilize the image Segmentation Technology of the potential structural information between image.
Summary of the invention
In order to effectively utilize the potential structural information between image, reduce the computation complexity of Iamge Segmentation process, promote the Target Segmentation effect in image, the present invention proposes a kind of multi-modality images dividing method based on Functional Mapping, the method comprises the following steps:
1, after acquisition comprises the image collection of target, following operation is carried out:
1) each Iamge Segmentation in set is become super-pixel block, represent to obtain multi-modality images by the super-pixel after different Feature Descriptors characterizes segmentation;
2) on multi-modality images, set up the figure based on super-pixel, and build corresponding Laplacian Matrix;
3) what characterize every width image about subtracts functional space, set up image between Functional Mapping;
4) the image Functional Mapping of often kind of mode is alignd with image clue, and introduce the consistance between implicit function maintenance Functional Mapping;
5) obtain Functional Mapping according to multi-modal consistency in mapping to express, by the segmentation function that combined optimization objective function computed image is corresponding, obtain image optimum segmentation and represent, complete Iamge Segmentation.
Further, described step 1) described in each Iamge Segmentation in set is become super-pixel block, represent to obtain multi-modality images by the super-pixel after different Feature Descriptors characterizes segmentation, specifically:
1) image establishing set to be associated by n width forms, and is designated as
every width image contains one or more target class, and the target class number of whole set is C;
2) pixel in image is regarded as the summit of figure, utilize figure dividing method that the image in set is divided into q zonule (as 100), q is positive integer, these zonules are made up of the pixel that value is close, be referred to as super-pixel, the block belonging to c class in the i-th width image is expressed as S
ic, wherein i={1,2 ..., n}, c={1,2 ..., C};
3) Feature Descriptor that m kind is different is utilized, as Scale invariant features transform (SIFT), local binary patterns (LBP), histogram of gradients (HOG) etc., each super-pixel in token image, thus obtain the multi-modal character representation of multi-faceted reflection image extrinsic information, as the i-th width image homography set
i.e. kth kind image feature descriptor corresponding kth kind mode
Further, described step 2) in the figure set up on multi-modality images based on super-pixel, and build corresponding Laplacian Matrix, specifically:
2.1) super-pixel of the q in often kind of image modalities is regarded as the summit of figure, build the super-pixel figure that respective vertices is formed by connecting entirely;
2.2) on the super-pixel figure of variant mode, Laplacian Matrix is built respectively
the weight matrix W that they are calculated by Gauss's weighted strategy obtains, i.e. L=D-W, and wherein D is pair of horns battle array, its diagonal entry be W each column element and.
Described step 3) in the linear of the every width image of sign about subtract functional space, set up image between Functional Mapping, specifically:
1) multi-modal Laplacian Matrix is calculated
eigenwert and proper vector, and get front p (p<q) individual proper vector Zhang Chengyue and subtract functional space
and characteristic of correspondence value forms diagonal matrix respectively in each mode of every width image
2) set every width Iamge Segmentation function as f
oicorresponding S
ic, the corresponding one group of base vector in search volume of this function
the p dimension space opened
and f
oilinear function combination is expressed as to the coefficient of the i-th width image
wherein B
ifor p before Laplacian Matrix proper vector is formed;
3) relation of reflection arbitrarily in pairs between image is mapped, as the subspace from the i-th width image by linear functional
to the subspace of jth width image
functional Mapping matrix
represent, i.e. subspace
in Function Mapping to subspace
in expression value can by calculating R
ijf obtains.
Described step 4) in the image Functional Mapping of often kind of mode is alignd with image clue, and introduce the consistance that implicit function keeps between Functional Mapping, specifically:
1) the corresponding different description operator of image clue, the image Functional Mapping of often kind of mode aligns with image clue and to realize by optimizing following formula, namely
Wherein, constant α >0, β >0, symbol ‖ ‖
1the L1 norm of representing matrix, symbol ‖ ‖
fthe Frobenius norm of representing matrix;
2) implicit function introduced is shared by input picture, and makes implicit function corresponding on the every width image of Functional Mapping energy efficient association by the consistent item of Functional Mapping, and each implicit function only appears in certain subset of image, and the implicit function z that the i-th width image is corresponding
i=[z
i1, z
i2..., z
il] ∈ 0,1} characterizes the relation between implicit function and image, and continuous variable Φ
i=[φ
i1, φ
i2..., φ
il] each implicit function on image is described;
The consistent item of Functional Mapping in previous step is expressed as
Wherein, constant γ >0, λ >0, symbol ‖ ‖
2the L2 norm of representing matrix, diag (z
i) representing pair of horns matrix, (i, j) ∈ Ε represents neighbour's set that image is right, as desirable 20 width neighbor image calculate.
Described step 5) in obtain Functional Mapping according to multi-modal consistency in mapping and express, calculate segmentation function corresponding to every width figure by combined optimization objective function, specifically:
1) calculate Functional Mapping according to the multi-modal consistency in mapping relation set up to express, namely
Wherein, variable Φ
iwith Φ
jbetween there is orthogonality constraint, adopt variable alternate optimization method to solve here, namely fix other Two Variables and optimize a remaining variable, variable z
ibe initialized as complete 1 vector, by successive ignition until function convergence, optimum Functional Mapping can be calculated and express R
ij;
2) take image pattern as figure
on summit, the weight between two summits is designated as
then the combined optimization goal expression of Iamge Segmentation function is
Wherein, constant ζ >0, c={1,2 ..., C}, symbol ()
trepresent vector or transpose of a matrix, subspace B
ikopened into by p proper vector before super-pixel figure Laplacian Matrix in the i-th width image corresponding kth kind mode, and different classes of segmentation function f
icmeet mutual exclusive restrict;
3) by solving the optimum solution of objective function in above-mentioned steps, the optimum segmentation function of the i-th width image can be obtained
can determine that belonging to other optimum segmentation block of c target class in image is expressed as accordingly
The present invention proposes the multi-modality images dividing method based on Functional Mapping, its advantage is: splitting formation super-pixel by carrying out figure to image original pixels, reducing computing cost; The sign content of the angle reflection image from different descriptor is represented by building multi-modal super-pixel; By about subtract functional space set up image between Functional Mapping, and utilize implicit function to keep its consistance, effectively establish the low-level feature of image and associating between high-level semantic, and then improve image segmentation, for such as the basis of compacting has been established in the application of the vision such as image recognition, target localization.
Accompanying drawing explanation
Fig. 1 is method flow diagram of the present invention.
Embodiment
With reference to accompanying drawing 1, further illustrate the present invention:
1, after acquisition comprises the image collection of target, following operation is carried out:
1) each Iamge Segmentation in set is become super-pixel block, represent to obtain multi-modality images by the super-pixel after different Feature Descriptors characterizes segmentation;
2) on multi-modality images, set up the figure based on super-pixel, and build corresponding Laplacian Matrix;
3) what characterize every width image about subtracts functional space, set up image between Functional Mapping;
4) the image Functional Mapping of often kind of mode is alignd with image clue, and introduce the consistance between implicit function maintenance Functional Mapping;
5) obtain Functional Mapping according to multi-modal consistency in mapping to express, by the segmentation function that combined optimization objective function computed image is corresponding, obtain image optimum segmentation and represent, complete Iamge Segmentation.
Step 1) described in by set in each Iamge Segmentation become super-pixel block, with different Feature Descriptors characterize segmentation after super-pixel represent to obtain multi-modality images, specifically:
1) image establishing set to be associated by n width forms, and is designated as
every width image contains one or more target class, and the target class number of whole set is C;
2) pixel in image is regarded as the summit of figure, utilize figure dividing method that the image in set is divided into q zonule (as 100), q is positive integer, these zonules are made up of the pixel that value is close, be referred to as super-pixel, the block belonging to c class in the i-th width image is expressed as S
ic, wherein i={1,2 ..., n}, c={1,2 ..., C};
3) Feature Descriptor that m kind is different is utilized, as Scale invariant features transform (SIFT), local binary patterns (LBP), histogram of gradients (HOG) etc., each super-pixel in token image, thus obtain the multi-modal character representation of multi-faceted reflection image extrinsic information, as the i-th width image homography set
i.e. kth kind image feature descriptor corresponding kth kind mode
Step 2) in the figure set up on multi-modality images based on super-pixel, and build corresponding Laplacian Matrix, specifically:
2.1) super-pixel of the q in often kind of image modalities is regarded as the summit of figure, build the super-pixel figure that respective vertices is formed by connecting entirely;
2.2) on the super-pixel figure of variant mode, Laplacian Matrix is built respectively
the weight matrix W that they are calculated by Gauss's weighted strategy obtains, i.e. L=D-W, wherein D to be diagonal entry be W column element and diagonal matrix.
Step 3) in the linear of the every width image of sign about subtract functional space, set up image between Functional Mapping, specifically:
1) multi-modal Laplacian Matrix is calculated
eigenwert and proper vector, and get front p (p<q) individual proper vector Zhang Chengyue and subtract functional space
and characteristic of correspondence value forms diagonal matrix respectively in each mode of every width image
2) set every width Iamge Segmentation function as f
oicorresponding S
ic, the corresponding one group of base vector in search volume of this function
the p dimension space opened
and f
oilinear function combination is expressed as to the coefficient of the i-th width image
wherein B
ifor p before Laplacian Matrix proper vector is formed;
3) relation of reflection arbitrarily in pairs between image is mapped, as the subspace from the i-th width image by linear functional
to the subspace of jth width image
functional Mapping matrix
represent, i.e. subspace
in Function Mapping to subspace
in expression value can by calculating R
ijf obtains.
Step 4) in the image Functional Mapping of often kind of mode is alignd with image clue, and introduce the consistance that implicit function keeps between Functional Mapping, specifically:
1) the corresponding different description operator of image clue, the image Functional Mapping of often kind of mode aligns with image clue and to realize by optimizing following formula, namely
Wherein, constant α >0, β >0, symbol ‖ ‖
1the L1 norm of representing matrix, symbol ‖ ‖
fthe Frobenius norm of representing matrix;
2) implicit function introduced is shared by input picture, and makes implicit function corresponding on the every width image of Functional Mapping energy efficient association by the consistent item of Functional Mapping, and each implicit function only appears in certain subset of image, and the implicit function z that the i-th width image is corresponding
i=[z
i1, z
i2..., z
il] ∈ 0,1} characterizes the relation between implicit function and image, and continuous variable Φ
i=[φ
i1, φ
i2..., φ
il] each implicit function on image is described;
The consistent item of Functional Mapping in previous step is expressed as
Wherein, constant γ >0, λ >0, symbol ‖ ‖
2the L2 norm of representing matrix, diag (z
i) representing pair of horns matrix, (i, j) ∈ Ε represents neighbour's set that image is right, as desirable 20 width neighbor image calculate.
Step 5) in obtain Functional Mapping according to multi-modal consistency in mapping and express, calculate segmentation function corresponding to every width figure by combined optimization objective function, specifically:
1) calculate Functional Mapping according to the multi-modal consistency in mapping relation set up to express, namely
Wherein, variable Φ
iwith Φ
jbetween there is orthogonality constraint, adopt variable alternate optimization method to solve here, namely fix other Two Variables and optimize a remaining variable, variable z
ibe initialized as complete 1 vector, by successive ignition until function convergence, optimum Functional Mapping can be calculated and express R
ij;
2) take image pattern as figure
on summit, the weight between two summits is designated as
then the combined optimization goal expression of Iamge Segmentation function is
Wherein, constant ζ >0, c={1,2 ..., C}, symbol ()
trepresent vector or transpose of a matrix, subspace B
ikopened into by p proper vector before super-pixel figure Laplacian Matrix in the i-th width image corresponding kth kind mode, and different classes of segmentation function f
icmeet mutual exclusive restrict;
3) by solving the optimum solution of objective function in above-mentioned steps, the optimum segmentation function of the i-th width image can be obtained
can determine that belonging to other optimum segmentation block of c target class in image is expressed as accordingly