CN105069774B - The Target Segmentation method of optimization is cut based on multi-instance learning and figure - Google Patents
The Target Segmentation method of optimization is cut based on multi-instance learning and figure Download PDFInfo
- Publication number
- CN105069774B CN105069774B CN201510375307.4A CN201510375307A CN105069774B CN 105069774 B CN105069774 B CN 105069774B CN 201510375307 A CN201510375307 A CN 201510375307A CN 105069774 B CN105069774 B CN 105069774B
- Authority
- CN
- China
- Prior art keywords
- matrix
- image
- brightness
- color
- pixel
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of Target Segmentation method that optimization is cut based on multi-instance learning and figure：Step 1：Notable model modeling is carried out using the method for multi-instance learning to training image, and the bag in test image and example are predicted using notable model, obtains the conspicuousness testing result of test image；Step 2：The conspicuousness testing result introducing figure of test image is cut into framework, the mark according to exemplary characteristics vector and example bag cuts framework to figure and optimized, and solution figure cuts the suboptimal solution of optimization, obtains the Accurate Segmentation of target.The present invention establishes conspicuousness detection model using the method for multi-instance learning, it is adapted for particular kind of image, and the result of conspicuousness detection is used to guide image split in the image partition method based on graph theory, model framework link is cut to figure to be optimized, and use Agglomerative Hierarchical Clustering Algorithm for Solving, enable segmentation result to conform better to the output of Semantic Aware, obtain accurate object segmentation result.
Description
Technical field
The invention belongs to image processing field, is related to a kind of image partition method, is specifically that one kind is based on multi-instance learning
The Target Segmentation method of optimization is cut with figure.
Background technology
Image object segmentation is an important research direction of computer vision field, while is also vision-based detection, tracking
With the important foundation of the application such as identification, it is split the quality of quality and largely affects the performance of whole vision system.
Recognize yet with the deep layer lacked to human visual system, image segmentation also becomes one of computer vision field simultaneously
Classic problem.The main contents of scene observed by human visual system can selectively pay attention to, and ignore other minor coverages.
This Selective Attention Mechanism of vision makes it possible efficient information processing, while has also inspired grinding for computer vision
The persons of studying carefully look for another way from the angle of attention mechanism, therefore the Image Segmentation Model with human visual system will be as image point
Cut the new study hotspot in one, field.
Detected from the conspicuousness of computer vision angle, it is broadly divided into bottom-up and top-down method.
Conspicuousness detection is to be based on non-supervisory model mostly at present, defined model be present and lacks learning ability, the calculating of significance
It can not reflect vision noticing mechanism well, and ask particular kind of image adaptability deficiency and robustness are poor etc.
Topic；And figure of the single use based on cost function cuts algorithm and carry out Target Segmentation, there is also computation complexity height, segmentation efficiency
The problem such as low with accuracy of local segmentation.
The content of the invention
For above-mentioned the shortcomings of the prior art, it is an object of the present invention to which proposing one kind is based on multi-instance learning
The Target Segmentation method of optimization is cut with figure, conspicuousness detection model is established so as to fit specific using the method for multi-instance learning
The image of species, and the result of conspicuousness detection is used to guide image split in the image partition method based on graph theory, it is right
Figure cuts all too many levels such as model framework and is optimized, and cuts the solution side of optimization as figure using Agglomerative Hierarchical Clustering algorithm
Method so that segmentation result can conform better to the output of Semantic Aware, obtain accurate object segmentation result.
To achieve these goals, the present invention, which adopts the following technical scheme that, is solved：
The Target Segmentation method of optimization is cut based on multi-instance learning and figure, is comprised the following steps：
Step 1：Notable model modeling is carried out using the method for multi-instance learning to training image, and utilizes notable model pair
Bag and example in test image are predicted, and obtain the conspicuousness testing result of test image；Specifically include：
Step 11, training image is pre-processed, and extracts gradient of image intensity feature and color gradient feature；
Step 12, multi-instance learning is incorporated into saliency detection, obtains the conspicuousness detection knot of test image
Fruit；
Step 2：The conspicuousness testing result introducing figure of test image is cut into framework, according to exemplary characteristics vector and example bag
Mark framework cut to figure optimized, solution figure cuts the suboptimal solution of optimization, obtains the Accurate Segmentation of target.
Further, training image is pre-processed in the step 11, and extracts brightness step feature and color ladder
Feature is spent, is specifically included：
Step 111, the conversion of color space is carried out to training image and its quantization of each component pre-processes, is normalized
Luminance component L and color component a, b afterwards；
Step 112, the brightness step of each pixel in luminance component L matrix is calculated；
Step 113, the color gradient of each pixel in color component a and color component b matrix is calculated respectively.
3rd, the Target Segmentation method of optimization is cut based on multi-instance learning and figure as claimed in claim 2, it is characterised in that
The step 111 is specific as follows：
First, training image is subjected to gamma correction, to realize the nonlinear adjustment to image color component, training schemed
As being changed by rgb color space to Lab color spaces；Again to luminance component L of the training image under Lab color spaces and two
Color component a, b are normalized, luminance component L and color component a, b after being normalized.
Further, the step 112 specifically includes step A-D：
A, the weight matrix Wights ＜ ＞ of 3 yardsticks are built；
B, the key map matrix Slice_map ＜ ＞ of 3 yardsticks are built；The key map matrix Slice_ of each yardstick
The weight matrix Wights ＜ ＞ that map ＜ ＞ correspond to yardstick have identical dimension；Choose 8 directions (0 °, 22.5 °, 45 °,
67.5 °, 90 °, 112.5 °, 135 °, 157.5 °) matrix is divided into 16 regions, the value of element and the region in each region
Numbering 0~15 it is identical；
C, by the weight matrix Wights ＜ ＞ of the corresponding yardsticks of each key map matrix Slice_map ＜ ＞
Element, which corresponds to be multiplied, obtains the matrix of corresponding yardstick, i.e. neighborhood gradient operator；
D, using neighborhood gradient operator, the brightness step of a pixel to be asked in luminance component L matrix is calculated；
Further, the step A is specific as follows：
The weight matrix Wights ＜ ＞ of 3 yardsticks are built respectively；Described weight matrix Wights ＜ ＞ be line number and
Columns is equal to 2r+1 square formation；Element non-zero i.e. 1 in weight matrix Wights ＜ ＞, the Elemental redistribution equal to 1 is with square formation
Central element (r+1, r+1) be the center of circle, using r as the disk of radius in the range of, form the inscribed circle of square formation, remaining element in square formation
It is 0；3 yardsticks are respectively r=3, r=5 and r=10.
Further, the step D is specific as follows：
1. for some yardstick, in the luminance component L obtained by step 111 matrix centered on a pixel to be asked,
Dot product is carried out by the neighborhood gradient operator and each luminance component in the range of neighborhood of pixel points to be asked of a certain yardstick, treated
Seek the matrix N eibor ＜ ＞ in the range of neighborhood of pixel points；The straight line of vertical direction (90 °) is chosen as line of demarcation, by neighborhood ladder
Disk in degree operator is divided into left semicircle and right semi-circle, and left semicircle includes the 0th sector to the 7th sector, and right semi-circle includes the 8th fan
Area to the 15th sector；Matrix N eibor ＜ ＞ element forms a histogram and carries out normalizing to it corresponding to each semicircle
Change, be designated as Slice_hist respectively1＜ ＞ and Slice_hist2＜ ＞；H1The histogram corresponding to the half-circle area of the left side is represented,
H2The histogram corresponding to the half-circle area of the right is represented, i is the bin of histogram value, is defined as [0,24], i.e. brightness model
Enclose；
2. calculating the difference between two normalization histograms by card side's distance shown in formula (1), that is, obtain a certain chi
The brightness step spent on the vertical direction of next pixel to be asked；
After brightness step on a certain yardstick vertical direction has been calculated, straight line conduct where other directions is chosen respectively
Line of demarcation, obtain the brightness step on the every other direction of a certain yardstick of pixel to be asked；Further according to the same modes of step D
The directive brightness step of institute on the pixel to be asked other yardsticks is calculated；When completion all yardsticks of pixel to be asked
After brightness step on all directions calculates, the final brightness step of the pixel to be asked is calculated by formula (2)：
f(x,y,r,n_ori；R=3,5,10；N_ori=1,2 ... 8)-＞ Brightness Gradient (x,
y) (2)
In formula, f is a mapping function, and (x, y) is any pixel to be asked, and r represents the yardstick chosen, and n_ori represents choosing
The direction taken；Brightness Gradient (x, y) are the final brightness step of pixel (x, y)；F correspondence rule is choosing
High-high brightness Grad of each direction in 3 yardsticks is selected as luminance gradient value in this direction, will be bright on 8 directions
Degree gradient sums to obtain the final brightness step of pixel (x, y).
Further, in the step 113, the calculating of color gradient is similar with the calculating of brightness step, the difference is that color
Color Gradient Features are the color gradient a and b for two color components；3 yardsticks chosen are respectively r=5, r=10 and r=
20；Therefore, the size of corresponding weight matrix and map reference matrix is respectively 11*11,21*21 and 41*41；Two colors point
The calculating of the color gradient of amount and brightness step use identical computational methods, obtain each waiting to ask in color component a and b matrix
The final color gradient of pixel.
Further, multi-instance learning is introduced into saliency in the step 12 to detect to obtain the aobvious of test image
Work property testing result, specifically includes step 121 and step 122：
Step 121, the brightness obtained using method described in step 11 and color gradient feature, with reference to multi-instance learning
EMDD algorithms realize the study to training set, obtain the conspicuousness detection model succeeded in school；
Step 122, test image is substituted into the conspicuousness detection model succeeded in school, obtains the conspicuousness detection of test image
As a result.
Further, described step 2 specifically comprises the following steps：
Step 21, the conspicuousness testing result of image step 1 obtained cuts the input of algorithm as figure, according to the aobvious of bag
Work property mark builds the weight function as shown in formula (3) with exemplary characteristics vector；And the figure obtained after the optimization as shown in formula (4) is cut
Cost function；
In formula (3), wijRepresent the visual signature similitude of i examples bag and j example bags corresponding region, Salien (i) with
Salien (j) represents the notable angle value after region i and region j normalization respectively, and σ is the sensitive ginseng of regulation visual signature difference
Number, value are 10~20；I weights similar to its own in region are 0；Similarity matrix W={ wijBe diagonal be 0 it is symmetrical
Matrix, and wij∈[0,1]；fi,fjRepresent i with distinguishing corresponding exemplary characteristics vector, the i.e. brightness step of image in j example bags
Feature synthesizes the mix vector Mixvector of 3-dimensional with color gradient characteristic vectori={ BrightnessGradienti,
ColorGradienti, then Sim (fi,fj)=‖ Mixvectori-Mixvectorj‖2.Figure represented by formula (4) is cut in framework,
D is N-dimensional diagonal matrix, element on its diagonalU={ U1,U2,...,Ui,...,Uj,...UNIt is segmentation shape
State vector, each component of a vector UiRepresent region i cutting state；The molecule of formula (4) is represented between region i and region j
Visual similarity, denominator represent the visual similarity in the i of region；
Step 22, the cutting state vector corresponding to R (U) minimum value characteristic value is solved, that is, obtains the most optimal sorting of image
Cut result.
Brief description of the drawings
Fig. 1 is the flow chart of the inventive method.
Fig. 2 is the comparative result figure that test image is split by multiple methods.Wherein, subgraph (a-1) to (a-4) is
Original test image, subgraph (b-1) to (b-4) be based on multiple dimensioned figure decompose spectrum partitioning algorithm segmentation result, subgraph
(c-1) to (c-4) directly to use the segmentation result of Agglomerative Hierarchical Clustering algorithm, subgraph (d-1) to (d-4) is the inventive method
Segmentation result.
Fig. 3 is disk or so subregion schematic diagram.
Fig. 4 is H1,H2Histogram schematic diagram.
Fig. 5 is to change disk line of demarcation direction schematic diagram.
The present invention is further explained with embodiment below in conjunction with accompanying drawing.
Embodiment
As shown in figure 1, the Target Segmentation method that optimization is cut based on multi-instance learning and figure of the present invention, is specifically included as follows
Step：
Step 1：Notable model modeling is carried out using the method for multi-instance learning to training image, and utilizes notable model pair
Bag and example in test image are predicted, and obtain the conspicuousness testing result of test image；
Step 2：The significance introducing figure of test image is cut into framework, the mark pair according to exemplary characteristics vector and example bag
Figure cuts framework and optimized, and the suboptimal solution of optimization is cut using Agglomerative Hierarchical Clustering Algorithm for Solving figure, obtains the Accurate Segmentation of target.
Further, described step 1 includes step 11 and step 12：
Step 11, training image is pre-processed, and extracts gradient of image intensity feature and color gradient feature；
Step 12, multi-instance learning is incorporated into saliency detection, obtains the conspicuousness detection knot of test image
Fruit.
Further, training image is pre-processed in the step 11, and extracts brightness step feature and color ladder
Feature is spent, specifically includes step 111~step 113：
Step 111, the conversion of color space is carried out to training image and its quantization of each component pre-processes, is normalized
Luminance component L and color component a, b afterwards；It is specific as follows：
First, training image is subjected to gamma correction, to realize the nonlinear adjustment to image color component, training schemed
As being changed by rgb color space to Lab color spaces；Again to luminance component L of the training image under Lab color spaces and two
Color component a, b are normalized, luminance component L and color component a, b after being normalized；
Step 112, the brightness step of each pixel in luminance component L matrix is calculated.Specifically include step A-D：
A, the weight matrix Wights ＜ ＞ of 3 yardsticks are built.It is specific as follows：
The weight matrix Wights ＜ ＞ of 3 yardsticks are built respectively；Described weight matrix Wights ＜ ＞ be line number and
Columns is equal to 2r+1 square formation；Element non-zero i.e. 1 in weight matrix Wights ＜ ＞, the Elemental redistribution equal to 1 is with square formation
Central element (r+1, r+1) be the center of circle, using r as the disk of radius in the range of, form the inscribed circle of square formation, remaining element in square formation
It is 0；In the present invention, when 3 yardsticks are respectively r=3, r=5 and r=10, weight matrix Wights ＜ ＞ corresponding to difference are such as
Under：
B, the key map matrix Slice_map ＜ ＞ of 3 yardsticks are built；The key map matrix Slice_ of each yardstick
The weight matrix Wights ＜ ＞ that map ＜ ＞ correspond to yardstick have identical dimension, i.e., each key map Slice_map ＜ ＞
Matrix is also the square formation that line number and columns are all 2r+1；Choose 8 directions (0 °, 22.5 °, 45 °, 67.5 °, 90 °, 112.5 °,
135 °, 157.5 °) matrix is divided into 16 regions, the value of element is identical with the numbering 0~15 in the region in each region；Build
Vertical key map matrix Slice_map ＜ ＞ purpose is to realize the fast positioning to subregion.In the present invention, 3 indexes
Map matrix Slice_map ＜ ＞ difference is as follows：
C, by the weight matrix Wights ＜ ＞ of the corresponding yardsticks of each key map matrix Slice_map ＜ ＞
Element, which corresponds to be multiplied, obtains the matrix of corresponding yardstick, i.e. neighborhood gradient operator.Neighborhood gradient operator under 3 yardsticks is such as
Under：
D, using neighborhood gradient operator, the brightness step of a pixel to be asked in luminance component L matrix is calculated.Specifically
It is as follows：
1. for some yardstick, in the luminance component L obtained by step 111 matrix centered on a pixel to be asked,
Dot product is carried out by the neighborhood gradient operator and each luminance component in the range of neighborhood of pixel points to be asked of a certain yardstick, treated
Seek the matrix N eibor ＜ ＞ in the range of neighborhood of pixel points；The straight line of vertical direction (90 °) is chosen as line of demarcation, by neighborhood ladder
Disk in degree operator is divided into left semicircle and right semi-circle, and left semicircle includes the 0th sector to the 7th sector, and right semi-circle includes the 8th fan
Area to the 15th sector；Matrix N eibor ＜ ＞ element forms a histogram and carries out normalizing to it corresponding to each semicircle
Change, be designated as Slice_hist respectively1＜ ＞ and Slice_hist2＜ ＞；As shown in Figure 4.H1Represent corresponding to the half-circle area of the left side
Histogram, H2The histogram corresponding to the half-circle area of the right is represented, i is the bin of histogram value, is defined as [0,24],
That is brightness range.
2. calculating the difference between two normalization histograms by card side's distance shown in formula (1), that is, obtain a certain chi
The brightness step spent on the vertical direction of next pixel to be asked；
After brightness step on a certain yardstick vertical direction has been calculated, as shown in figure 5, choosing other direction institutes respectively
In straight line as line of demarcation, the brightness step on the every other direction of a certain yardstick of pixel to be asked is obtained；Further according to step D
The directive brightness step of institute on the pixel to be asked other yardsticks is calculated in same mode.When the completion pixel to be asked
After brightness step on all directions of yardstick of point calculates, the final brightness of the pixel to be asked is calculated by formula (2)
Gradient：
f(x,y,r,n_ori；R=3,5,10；N_ori=1,2 ... 8)-＞ Brightness Gradient (x,
y) (2)
In formula, f is a mapping function, and (x, y) is any pixel to be asked, and r represents the yardstick chosen, and n_ori represents choosing
The direction taken；Brightness Gradient (x, y) are the final brightness step of pixel (x, y)；F correspondence rule is choosing
High-high brightness Grad of each direction in 3 yardsticks is selected as luminance gradient value in this direction, will be bright on 8 directions
Degree gradient sums to obtain the final brightness step of pixel (x, y)；
Step 113, the color gradient of each pixel in color component a and color component b matrix is calculated respectively.Tool
Body is as follows：
The calculating of color gradient is similar with the calculating of brightness step, the difference is that color gradient is characterized in being directed to two colors
Color component a and b under the color gradient of component, i.e. Lab color spaces；It is with the calculating difference of brightness step, selects
3 yardsticks taken are respectively r=5, r=10 and r=20；Therefore, the size of corresponding weight matrix and map reference matrix point
Wei not 11*11,21*21 and 41*41；The calculating of the color gradient of two color components and brightness step use identical calculating side
Method, obtain in color component a and b matrix each final color gradient of pixel to be asked.
Further, multi-instance learning is introduced into saliency in step 12 to detect to obtain the conspicuousness of test image
Testing result, specifically include step 121 and step 122：
Step 121, the brightness obtained using method described in step 11 and color gradient feature, with reference to multi-instance learning
EMDD algorithms realize the study to training set, obtain the conspicuousness detection model succeeded in school.Comprise the following steps that：
Region segmentation, the minimum pixel number for including each region are carried out to training image using oversubscription segmentation method first
For 200；Each region is taken as a bag, and to each region progress stochastical sampling, the pixel in the region being sampled is taken as
Example, corresponding brightness step feature is extracted with color gradient characteristic vector as sampling instances characteristic vector；Shown according to sampling
Example characteristic vector, the training of grader is carried out using multi-instance learning method EMDD algorithms, obtains the conspicuousness detection succeeded in school
Model；
Step 122, test image is substituted into the conspicuousness detection model succeeded in school, obtains the conspicuousness detection of test image
As a result.
To each width test image, test image is pre-processed using with step 11 identical process, obtains brightness
Gradient Features and color gradient feature；Then region segmentation is carried out to test image using oversubscription segmentation method, wraps each region
The minimum pixel number contained is 200；Each region as a bag and is subjected to stochastical sampling, the area being sampled to each region
Pixel is taken as example in domain, extracts corresponding brightness step feature with color gradient characteristic vector as sampling instances Characteristic Vectors
Amount, conspicuousness detection model succeed in school obtained using step 121, obtain that significant exemplary characteristics vector each wraps shows
Work property, so as to obtain the conspicuousness testing result of test image.
Further, described step 2 specifically comprises the following steps：
Step 21, the conspicuousness testing result of image step 1 obtained cuts the input of algorithm as figure, according to the aobvious of bag
Work property mark builds the weight function as shown in formula (3) with exemplary characteristics vector；And the figure obtained after the optimization as shown in formula (4) is cut
Cost function；
In formula (3), wijRepresent the visual signature similitude of i examples bag and j example bags corresponding region, Salien (i) with
Salien (j) represents the notable angle value after region i and region j normalization respectively, and σ is the sensitive ginseng of regulation visual signature difference
Number, value are 10~20；I weights similar to its own in region are 0；Similarity matrix W={ wijBe diagonal be 0 it is symmetrical
Matrix, and wij∈[0,1]；fi,fjRepresent i with distinguishing corresponding exemplary characteristics vector, the i.e. brightness step of image in j example bags
Feature synthesizes the mix vector Mixvector of 3-dimensional with color gradient characteristic vectori={ BrightnessGradienti,
ColorGradienti, then Sim (fi,fj)=‖ Mixvectori-Mixvectorj‖2.Figure represented by formula (4) is cut in framework,
D is N-dimensional diagonal matrix, element on its diagonalU={ U1,U2,...,Ui,...,Uj,...UNIt is segmentation shape
State vector, each component of a vector UiRepresent region i cutting state；The molecule of formula (4) is represented between region i and region j
Visual similarity, denominator represent the visual similarity in the i of region；
Step 22, using Agglomerative Hierarchical Clustering algorithm, solve cutting state corresponding to R (U) minimum value characteristic value to
Amount, that is, obtain the optimum segmentation result of image.
Wherein, above-mentioned Agglomerative Hierarchical Clustering algorithm refers to the step 2 for the method that number of patent application is 201210257591.1
With step 3.
Verification experimental verification
To verify the validity of the inventive method, using the database of Achanta et al. foundation, 300 therein are chosen
For training image, remaining 700 are that test image carries out proof of algorithm.Part of test results is enumerated as shown in Fig. 2 Fig. 2 distinguishes
Giving the segmentation result to test image, number of patent application based on the spectrum partitioning algorithm that multiple dimensioned figure decomposes is
201210257591.1 method to the segmentation result of test image, and the segmentation result using the inventive method.Illustrate such as
Under：
Subgraph (a-1) to (a-4) is original image in Fig. 2, and subgraph (b-1) to (b-4) is to be decomposed based on multiple dimensioned figure
The segmentation result of partitioning algorithm is composed, subgraph (c-1) to (c-4) is the method that number of patent application is 201210257591.1, subgraph
(d-1) to (d-4) is the inventive method.It can be drawn by experimental result, when background is relative complex, be decomposed based on multiple dimensioned figure
Spectrum partitioning algorithm serious mistake point and the imperfect phenomenon of Target Segmentation be present, and number of patent application is 201210257591.1
Method and the inventive method all have preferable segmentation result；When background is fairly simple and is differed greatly with target signature, such as
Original image (a-1) and (a-2), three kinds of methods can be partitioned into more complete target；But in target and background border mistake
Cross slowly and in the case that difference is minimum, such as original image (a-3) and (a-4), three kinds of methods all have different degrees of target
Split imperfect, but method and the number of patent application of the present invention is that the segmentation effect of 201210257591.1 method wants more preferable one
A bit, and method of the invention is dividedly more fine in the intersection of the minimum target of difference and background, can obtain notable mesh
The more accurate segmentation result of mark.Number of patent application is that the input picture of 201210257591.1 method is original image, slightly
The object of change is since Pixel-level, although pixel-level image is more fine, the consideration of amount of calculation is in, directly using bright
The figure segmentation method of degree and color character only considered gray difference when weight function defines, and the inventive method combines more examples
Learning method can obtain the marking area mark in image quickly, and the exemplary characteristics vector in each example bag contains reflection
The bottom visual signature of target information and the feature on the middle and senior level of objective contour, at the beginning of roughening, just consider comprehensive spy of image
Levy and provide accurately segmentation foundation for subsequent treatment, therefore when target is slow with background border transition and difference is minimum
Situation, it can still obtain preferable segmentation result.For most test image, the level iterations of the inventive method
Less than the iterations for the method that number of patent application is 201210257591.1, operand and time complexity are greatly reduced.
Claims (8)
- A kind of 1. Target Segmentation method that optimization is cut based on multi-instance learning and figure, it is characterised in that comprise the following steps：Step 1：Notable model modeling is carried out using the method for multi-instance learning to training image, and using notable model to test Bag and example in image are predicted, and obtain the conspicuousness testing result of test image；Specifically include：Step 11, training image is pre-processed, and extracts gradient of image intensity feature and color gradient feature；Step 12, multi-instance learning is incorporated into saliency detection, obtains the conspicuousness testing result of test image；Step 2：The conspicuousness testing result introducing figure of test image is cut into framework, the mark according to exemplary characteristics vector and example bag Note is cut framework to figure and optimized, and solution figure cuts the suboptimal solution of optimization, obtains the Accurate Segmentation of target；Described step 2 specifically comprises the following steps：Step 21, the conspicuousness testing result of image step 1 obtained cuts the input of algorithm as figure, the conspicuousness according to bag Mark builds the weight function as shown in formula (3) with exemplary characteristics vector；And the figure obtained after the optimization as shown in formula (4) cuts cost Function；In formula (3), wijRepresent i examples bag and the visual signature similitude of j example bags corresponding region, Salien (i) and Salien (j) the notable angle value after region i and region j normalization is represented respectively, δ is the sensitive parameter of regulation visual signature difference, value For 10~20；I weights similar to its own in region are 0；Similarity matrix W={ wijIt is the symmetrical matrix that diagonal is 0, and wij∈[0,1]；fi,fjRepresent i with j example bags distinguish corresponding exemplary characteristics vector, i.e., the brightness step feature of image and Color gradient characteristic vector synthesizes the mix vector Mixvector of 3-dimensionali={ BrightnessGradienti, ColorGradienti, then Sim (fi,fj)=| | Mixvectori-Mixvectorj||2；Figure represented by formula (4) cuts framework In, D is N-dimensional diagonal matrix, element on its diagonalU={ U1,U2,...,Ui,...,Uj,...UNIt is segmentation State vector,Each component of a vector UiRepresent region i cutting state；The molecule of formula (4) is represented between region i and region jVisual similarity, denominator represent region i in visual similarity；Step 22, the cutting state vector corresponding to R (U) minimum value characteristic value is solved, that is, obtains the optimum segmentation knot of image Fruit.
- 2. the Target Segmentation method of optimization is cut based on multi-instance learning and figure as claimed in claim 1, it is characterised in that described Training image is pre-processed in step 11, and extracts brightness step feature and color gradient feature, is specifically included：Step 111, the conversion of color space is carried out to training image and its quantization of each component pre-processes, after being normalized Luminance component L and color component a, b；Step 112, the brightness step of each pixel in luminance component L matrix is calculated；Step 113, the color gradient of each pixel in color component a and color component b matrix is calculated respectively.
- 3. the Target Segmentation method of optimization is cut based on multi-instance learning and figure as claimed in claim 2, it is characterised in that described Step 111 is specific as follows：First, training image is subjected to gamma correction, to realize to the nonlinear adjustment of image color component, by training image by Rgb color space is changed to Lab color spaces；Luminance component L and two colors to training image under Lab color spaces again Component a, b are normalized, luminance component L and color component a, b after being normalized.
- 4. the Target Segmentation method of optimization is cut based on multi-instance learning and figure as claimed in claim 2, it is characterised in that described Step 112 specifically includes step A-D：A, the weight matrix Wights ＜ ＞ of 3 yardsticks are built；B, the key map matrix Slice_map ＜ ＞ of 3 yardsticks are built；The key map matrix Slice_map of each yardstick The weight matrix Wights ＜ ＞ that ＜ ＞ correspond to yardstick have identical dimension；Choose 8 directions (0 °, 22.5 °, 45 °, 67.5 °, 90 °, 112.5 °, 135 °, 157.5 °) matrix is divided into 16 regions, the value of element and the region in each region Numbering 0~15 it is identical；C, by the element in the weight matrix Wights ＜ ＞ of the corresponding yardsticks of each key map matrix Slice_map ＜ ＞ Correspond to be multiplied and obtain the matrix of corresponding yardstick, i.e. neighborhood gradient operator；D, using neighborhood gradient operator, the brightness step of a pixel to be asked in luminance component L matrix is calculated.
- 5. the Target Segmentation method of optimization is cut based on multi-instance learning and figure as claimed in claim 4, it is characterised in that described Step A is specific as follows：The weight matrix Wights ＜ ＞ of 3 yardsticks are built respectively；Described weight matrix Wights ＜ ＞ are line number and columns It is equal to 2r+1 square formation；Element non-zero i.e. 1 in weight matrix Wights ＜ ＞, the Elemental redistribution equal to 1 is with square formation center Element (r+1, r+1) be the center of circle, using r as the disk of radius in the range of, form the inscribed circle of square formation, remaining element is in square formation 0；3 yardsticks are respectively r=3, r=5 and r=10.
- 6. the Target Segmentation method of optimization is cut based on multi-instance learning and figure as claimed in claim 4, it is characterised in that described Step D is specific as follows：1. for some yardstick, in the luminance component L obtained by step 111 matrix centered on a pixel to be asked, pass through The neighborhood gradient operator of a certain yardstick carries out dot product with each luminance component in the range of neighborhood of pixel points to be asked, and obtains picture to be asked Matrix N eibor ＜ ＞ in vegetarian refreshments contiguous range；The straight line of vertical direction (90 °) is chosen as line of demarcation, neighborhood gradient is calculated Disk in son is divided into left semicircle and right semi-circle, and left semicircle includes the 0th sector to the 7th sector, and right semi-circle arrives including the 8th sector 15th sector；Matrix N eibor ＜ ＞ element forms a histogram and it is normalized corresponding to each semicircle, point Slice_hist is not designated as it1＜ ＞ and Slice_hist2＜ ＞；H1Represent the histogram corresponding to the half-circle area of the left side, H2Represent Histogram corresponding to the half-circle area of the right, i are the bin of histogram value, are defined as [0,24], i.e. brightness range；2. calculating the difference between two normalization histograms by card side's distance shown in formula (1), that is, obtain under a certain yardstick Brightness step on the vertical direction of one pixel to be asked；After brightness step on a certain yardstick vertical direction has been calculated, straight line is as boundary where choosing other directions respectively Line, obtain the brightness step on the every other direction of a certain yardstick of pixel to be asked；Calculated further according to the same modes of step D Obtain the directive brightness step of institute on the pixel to be asked other yardsticks；Own when completing all yardsticks of pixel to be asked After brightness step on direction calculates, the final brightness step of the pixel to be asked is calculated by formula (2)：f(x,y,r,n_ori；R=3,5,10；N_ori=1,2 ... 8)-＞ Brightness Gradient (x, y) (2)In formula, f is a mapping function, and (x, y) is any pixel to be asked, and r represents the yardstick chosen, and n_ori represents what is chosen Direction；Brightness Gradient (x, y) are the final brightness step of pixel (x, y)；F correspondence rule is every for selection High-high brightness Grad of the individual direction in 3 yardsticks is as luminance gradient value in this direction, by the brightness ladder on 8 directions Degree summation obtains the final brightness step of pixel (x, y).
- 7. the Target Segmentation method of optimization is cut based on multi-instance learning and figure as claimed in claim 2, it is characterised in that institute State in step 113, the calculating of color gradient is similar with the calculating of brightness step, the difference is that color gradient is characterized in being directed to two The color gradient a and b of color component；3 yardsticks chosen are respectively r=5, r=10 and r=20；Therefore, corresponding weights square Battle array and the size of map reference matrix are respectively 11*11,21*21 and 41*41；The calculating of the color gradient of two color components and Brightness step uses identical computational methods, obtains the final color ladder of each pixel to be asked in color component a and b matrix Degree.
- 8. the Target Segmentation method of optimization is cut based on multi-instance learning and figure as claimed in claim 1, it is characterised in that described Multi-instance learning is introduced into saliency in step 12 to detect to obtain the conspicuousness testing result of test image, specifically included Step 121 and step 122：Step 121, the brightness obtained using method described in step 11 and color gradient feature, calculated with reference to multi-instance learning EMDD Method realizes the study to training set, obtains the conspicuousness detection model succeeded in school；Step 122, test image is substituted into the conspicuousness detection model succeeded in school, obtains the conspicuousness detection knot of test image Fruit.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510375307.4A CN105069774B (en) | 2015-06-30 | 2015-06-30 | The Target Segmentation method of optimization is cut based on multi-instance learning and figure |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510375307.4A CN105069774B (en) | 2015-06-30 | 2015-06-30 | The Target Segmentation method of optimization is cut based on multi-instance learning and figure |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105069774A CN105069774A (en) | 2015-11-18 |
| CN105069774B true CN105069774B (en) | 2017-11-10 |
Family
ID=54499132
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510375307.4A Active CN105069774B (en) | 2015-06-30 | 2015-06-30 | The Target Segmentation method of optimization is cut based on multi-instance learning and figure |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105069774B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110111340A (en) * | 2019-04-28 | 2019-08-09 | 南开大学 | The Weakly supervised example dividing method cut based on multichannel |
Families Citing this family (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105760886B (en) * | 2016-02-23 | 2019-04-12 | 北京联合大学 | A kind of more object segmentation methods of image scene based on target identification and conspicuousness detection |
| CN105894517B (en) * | 2016-04-22 | 2019-05-07 | 北京理工大学 | CT image liver segmentation method and system based on feature learning |
| CN105956563B (en) * | 2016-05-06 | 2019-04-16 | 西安工程大学 | The method for carrying out face mark in news image based on multi-instance learning |
| CN109643363B (en) | 2016-06-15 | 2023-06-09 | 诺基亚技术有限公司 | Method, system and device for feature extraction and object detection |
| CN108625844B (en) * | 2017-03-17 | 2024-06-25 | 中国石油化工集团有限公司 | Gamma calibration and testing device |
| CN110555122B (en) * | 2019-07-19 | 2022-12-23 | 中国地质大学（武汉） | Building plan wall vectorization method based on segmented rectangles |
| CN112164068B (en) * | 2020-07-17 | 2023-07-28 | 中国人民解放军63892部队 | Self-adaptive extraction method for warning radar PPI image target and interference area |
| CN113159026A (en) * | 2021-03-31 | 2021-07-23 | 北京百度网讯科技有限公司 | Image processing method, image processing apparatus, electronic device, and medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102509084A (en) * | 2011-11-18 | 2012-06-20 | 中国科学院自动化研究所 | Multi-examples-learning-based method for identifying horror video scene |
| CN103198321A (en) * | 2013-03-27 | 2013-07-10 | 中国科学院苏州生物医学工程技术研究所 | Retina image processing method and device |
| CN103996189A (en) * | 2014-05-05 | 2014-08-20 | 小米科技有限责任公司 | Image segmentation method and device |
| CN104134217A (en) * | 2014-07-29 | 2014-11-05 | 中国科学院自动化研究所 | Video salient object segmentation method based on super voxel graph cut |
| CN104217225A (en) * | 2014-09-02 | 2014-12-17 | 中国科学院自动化研究所 | A visual target detection and labeling method |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7672987B2 (en) * | 2005-05-25 | 2010-03-02 | Siemens Corporate Research, Inc. | System and method for integration of medical information |
| US8249366B2 (en) * | 2008-06-16 | 2012-08-21 | Microsoft Corporation | Multi-label multi-instance learning for image classification |
-
2015
- 2015-06-30 CN CN201510375307.4A patent/CN105069774B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102509084A (en) * | 2011-11-18 | 2012-06-20 | 中国科学院自动化研究所 | Multi-examples-learning-based method for identifying horror video scene |
| CN103198321A (en) * | 2013-03-27 | 2013-07-10 | 中国科学院苏州生物医学工程技术研究所 | Retina image processing method and device |
| CN103996189A (en) * | 2014-05-05 | 2014-08-20 | 小米科技有限责任公司 | Image segmentation method and device |
| CN104134217A (en) * | 2014-07-29 | 2014-11-05 | 中国科学院自动化研究所 | Video salient object segmentation method based on super voxel graph cut |
| CN104217225A (en) * | 2014-09-02 | 2014-12-17 | 中国科学院自动化研究所 | A visual target detection and labeling method |
Non-Patent Citations (4)
| Title |
|---|
| Global contrast based salient region detection;Ming-Ming Cheng;《Computer Vision & Pattern Recognition》;20150331;第37卷(第3期);第409-416页 * |
| 基于多特征的图像轮廓检测算法研究;许可;《中国优秀硕士学位论文全文数据库信息科技辑》;20150115;第3.1、3.3、3.4节 * |
| 基于多示例学习的图像显著性分析;谢延涛;《中国优秀硕士学位论文全文数据库信息科技辑》;20150215;第3.2节、图3-2、图3-10 * |
| 基于视觉显著模型和图割的图像分割方法研究;刘红才;《中国优秀硕士学位论文全文数据库信息科技辑》;20140615;第3.1节，图3.1 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110111340A (en) * | 2019-04-28 | 2019-08-09 | 南开大学 | The Weakly supervised example dividing method cut based on multichannel |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105069774A (en) | 2015-11-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105069774B (en) | The Target Segmentation method of optimization is cut based on multi-instance learning and figure | |
| CN108830188B (en) | Vehicle detection method based on deep learning | |
| CN104835175B (en) | Object detection method in a kind of nuclear environment of view-based access control model attention mechanism | |
| CN111160269A (en) | Face key point detection method and device | |
| CN107679531A (en) | Licence plate recognition method, device, equipment and storage medium based on deep learning | |
| Ge et al. | Enhanced subpixel mapping with spatial distribution patterns of geographical objects | |
| CN109711288A (en) | Remote sensing ship detecting method based on feature pyramid and distance restraint FCN | |
| CN109344874A (en) | A kind of automatic chromosome analysis method and system based on deep learning | |
| CN106503739A (en) | The target in hyperspectral remotely sensed image svm classifier method and system of combined spectral and textural characteristics | |
| CN105005989B (en) | A kind of vehicle target dividing method under weak contrast | |
| CN104408449B (en) | Intelligent mobile terminal scene literal processing method | |
| CN103218605B (en) | A kind of fast human-eye positioning method based on integral projection and rim detection | |
| CN105956560A (en) | Vehicle model identification method based on pooling multi-scale depth convolution characteristics | |
| CN104200228B (en) | Recognizing method and system for safety belt | |
| CN106778687A (en) | Method for viewing points detecting based on local evaluation and global optimization | |
| CN108268865A (en) | Licence plate recognition method and system under a kind of natural scene based on concatenated convolutional network | |
| CN103186904A (en) | Method and device for extracting picture contours | |
| CN103020649A (en) | Forest type identification method based on texture information | |
| CN106780582A (en) | Based on the image significance detection method that textural characteristics and color characteristic are merged | |
| CN112784869A (en) | Fine-grained image identification method based on attention perception and counterstudy | |
| CN104282008A (en) | Method for performing texture segmentation on image and device thereof | |
| CN105404868A (en) | Interaction platform based method for rapidly detecting text in complex background | |
| CN108021920A (en) | A kind of method that image object collaboration is found | |
| CN105069403B (en) | A kind of three-dimensional human ear identification based on block statistics feature and the classification of dictionary learning rarefaction representation | |
| CN103544488A (en) | Face recognition method and device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |