CN108491883A - A kind of conspicuousness inspection optimization method based on condition random field - Google Patents

A kind of conspicuousness inspection optimization method based on condition random field Download PDF

Info

Publication number
CN108491883A
CN108491883A CN201810256988.6A CN201810256988A CN108491883A CN 108491883 A CN108491883 A CN 108491883A CN 201810256988 A CN201810256988 A CN 201810256988A CN 108491883 A CN108491883 A CN 108491883A
Authority
CN
China
Prior art keywords
image
cluster
image cluster
random field
input picture
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
Application number
CN201810256988.6A
Other languages
Chinese (zh)
Other versions
CN108491883B (en
Inventor
牛玉贞
林文奇
柯逍
陈俊豪
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Fuzhou University
Original Assignee
Fuzhou University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Fuzhou University filed Critical Fuzhou University
Priority to CN201810256988.6A priority Critical patent/CN108491883B/en
Publication of CN108491883A publication Critical patent/CN108491883A/en
Application granted granted Critical
Publication of CN108491883B publication Critical patent/CN108491883B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • 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
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Image Analysis (AREA)

Abstract

The conspicuousness inspection optimization method based on condition random field that the present invention relates to a kind of, which is characterized in that include the following steps：Step S1：Extract the global depth convolution feature of each image in input picture set；Step S2：According to the similitude between image two-by-two in global depth convolution feature calculation input picture set；Step S3：K means clusters are carried out to input picture set according to the similitude between image, form k mutually independent image clusters；Step S4：The full condition of contact random field optimized parameter of each image cluster is calculated using trellis search method；Step S5：For new input picture, the image cluster belonging to it is judged, the Saliency maps of the new input picture are optimized using the full condition of contact random field optimized parameter of affiliated image cluster.This method is suitable for the optimization of a variety of conspicuousness detection algorithms, and effect of optimization is good.

Description

A kind of conspicuousness inspection optimization method based on condition random field
Technical field
It is especially a kind of based on condition random field the present invention relates to image and video processing and computer vision field Conspicuousness inspection optimization method.
Background technology
Saliency detection algorithm extracts the place of most arresting in image, this partial information is in practical applications The more other regions of importance compare bigger.Conspicuousness detection algorithm is widely used in including image segmentation, image classification, figure As in the practical applications such as retrieval.Therefore, a large amount of scholars study saliency detection algorithm and constantly propose newest algorithm. Margolin et al. proposes a kind of notable based on Principal Component Analysis (Principal Component Analysis, PCA) Property detection algorithm, which calculates pattern differentials and detects Saliency maps.Wei et al. proposes a kind of based on geodesic curve The conspicuousness detection algorithm of (Geodesic saliency, GS), the algorithm pay close attention to the background prior information of image.Kim et al. is carried Go out a kind of conspicuousness detection calculation based on higher-dimension color conversion (High-dimensional Color Transform, HDCT) The RGB color figure of low-dimensional is mapped to the color feature space of higher-dimension, and searches an optimal cie system of color representation cie in higher dimensional space by method Linear combination is counted to define salient region.But existing saliency detection algorithm is compared with the standard picture manually marked There is certain defect, for example PCA algorithms pay attention to extract the edge of notable object, it is virtually impossible to detect in notable object Between part.GS algorithms can only extract the subregion of notable object, and the background of a part is also detected as significantly by HDCT algorithms Sex object.
Invention content
The conspicuousness inspection optimization method based on condition random field that the purpose of the present invention is to provide a kind of, this method are applicable in In the optimization of a variety of conspicuousness detection algorithms, and effect of optimization is good.
To achieve the above object, the technical solution adopted by the present invention is：A kind of conspicuousness detection based on condition random field Optimization method includes the following steps：
Step S1：Extract the global depth convolution feature of each image in input picture set；
Step S2：According to the similitude between image two-by-two in global depth convolution feature calculation input picture set；
Step S3：K-means clusters are carried out to input picture set according to the similitude between image, form k mutually Independent image cluster；
Step S4：The full condition of contact random field optimized parameter of each image cluster is calculated using trellis search method；
Step S5：For new input picture, the image cluster belonging to it is judged, using the full condition of contact of affiliated image cluster Random field optimized parameter optimizes the Saliency maps of the new input picture.
Further, in the step S1, the global depth convolution feature of image is extracted, is included the following steps：
Step S11：Extract image partial-depth convolution feature, this feature by image recognition depth network last layer Convolutional layer generates, for the image I of input, the partial-depth convolution feature f of output d dimensions t × t；
Step S12：The aggregate weight of partial-depth convolution feature is calculated, calculation formula is as follows：
Wherein, α (x, y) indicates that the aggregate weight of pixel (x, y) in image I, σ take between picture centre and nearest boundary The 1/3 of distance；
Step S13：Weighting polymerization partial-depth convolution feature, obtains preliminary global characteristics, calculation formula is as follows：
Wherein,Indicate the preliminary global characteristics after image I polymerizations, dimension is the width that d, W and H indicate image respectively And height, f (x, y) indicate the partial-depth convolution feature of pixel (x, y) in image I；
Step S14：It is rightL2 standardization processings are carried out, calculation formula is as follows：
Wherein,Indicate the global characteristics after image I standardization, dimension is d；
Step S15：It willPCA dimensionality reductions and whitening processing are carried out, the global depth convolution feature of n dimensions is obtainedn< D, calculation formula are as follows：
Wherein, MPCAIt is the PCA matrixes of a n × d, viIt is association singular value, i ∈ { 1,2 ..., n }, diag (v1, v2,...,vn) indicate diagonal matrix；
Step S16：It is rightL2 standardization processings are carried out, global depth convolution feature to the end is obtainedIt calculates public Formula is as follows：
Further, in the step S2, the similitude between image uses corresponding global depth convolution feature two-by-two Between dot product calculate, calculation formula is as follows：
Wherein,<>Indicate point multiplication operation, S (Ip,Iq) indicate image IpAnd IqBetween similitude, the value of similitude is bigger Indicate image IpAnd IqIt is more similar,Image I is indicated respectivelyp、IqGlobal depth convolution feature.
Further, in the step S3, it is poly- that K-means is carried out to input picture set according to the similitude between image Class includes the following steps：
Step S31：The center that k images are image cluster is randomly selected, is then calculated by solving following optimization problem Other image IiAffiliated image cluster ci：
Wherein, UrIt is the center of image cluster, r=1,2 ..., k, S (Ii,Ur) it is image IiWith the center U of image clusterrBetween Similitude,Expression takes and image IiThe corresponding image cluster in center of the maximum image cluster of similitude, then It is assigned to ci, to image IiBelong to image cluster ci；
Step S32：The center of each image cluster takes the mean value of all images in the image cluster, forms new center, then It is clustered again by the step S31 methods for solving optimization problem；
Step S33：Continuous iterative step S32 obtains k image cluster { C until the center of each image cluster no longer changesr | r=1 ..., k }.
Further, in the step S4, the full condition of contact that each image cluster is calculated using trellis search method is random Field optimized parameter, includes the following steps：
Step S41：For every image in each image cluster, Saliency maps are generated using conspicuousness detection algorithm；
Step S42：For every image in each image cluster, it is random that full condition of contact is traversed using trellis search method The parameter ω of the energy function of field1、ω2、σα、σβ、σγAll valued combinations；The calculation formula of energy function is as follows：
Wherein, L indicates the notable label of whole image, L ∈ { 0,1 }W×H, { 0,1 }W×HIndicate taking for W × H a 0 or 1 Value；E (L) indicates the corresponding energy functions of notable label L, liAnd ljThe notable mark of pixel i and j in notable label L are indicated respectively Label, value are that 0 expression is not notable, and value is that 1 expression is notable；P(li) be pixel i be notable label liProbability, take P (1)= δi, P (0)=1- δi, δiIt is the significance value of pixel i in Saliency maps；Energy function includes two parts：First partFor unit potential function, remaining part is pairs of potential function, including two cores, and first core is to be based on pixel The bilateral core of distance and color distortion, position is close and the similar similar significance value of pixel distribution of color, the core use Parameter ω1Control, pixel distance and color distortion use parameter σαAnd σβControl；Second core only depends on pixel distance, It aims at and improves isolated point small in former Saliency maps, using parameter ω2And σγControl；μ(li,lj) it is decision function, work as li =ljWhen, value 0, otherwise, value 1；piAnd pjThe pixel position of pixel i and j are indicated respectively；riAnd rjIt indicates respectively The color value of pixel i and j；The parameter ω1、ω2、σα、σβ、σγValue range be set of integers；
For the parameter ω of every image1、ω2、σα、σβ、σγEach parameter value combination, complete calculate condition of contact with The minimization problem of the energy function on airport：
Expression takes the minimum corresponding notable label L of energy function E (L), is then assigned to significantly Label L *；
Each parameter value combination for every image, each picture in image is calculated according to the notable label L * solved The posterior probability of vegetarian refreshments, to generate the Saliency maps after optimization；
Step S43：Using this conspicuousness check and evaluation standard of accuracy rate recall rate area under the curve PR-AUC as most The basis for estimation of excellent parameter value combination calculates and owns in the lower image cluster of each parameter value combination for each image cluster The average PR-AUC values of Saliency maps after image optimization take the highest parameter value combination of PR-AUC values as the complete of the image cluster Condition of contact random field optimized parameter.
Further, in the step S5, the image cluster belonging to new input picture is judged, using the complete of affiliated image cluster Condition of contact random field optimized parameter optimizes the Saliency maps of the new input picture, includes the following steps：
Step S51：New input picture I is generated using with identical conspicuousness detection algorithm in step S41eConspicuousness Figure；
Step S52：By solving following optimization problem calculating input image IeAffiliated image cluster ce：
Wherein, UrIt is the center for the image cluster that step S3 is obtained, S (Ie,Ur) it is input picture IeWith the center U of image clusterr Between similitude,Expression takes and input picture IeThe corresponding figure in center of the maximum image cluster of similitude As cluster, it is then assigned to ce, to input picture IeBelong to image cluster ce；
Step S53：Using image cluster ceFull condition of contact random field optimized parameter according to the calculation formula in step S42 To input picture IeSaliency maps optimize, generate optimization after Saliency maps.
Compared to the prior art, the beneficial effects of the invention are as follows：Using the maximum matching method connected entirely, using connecting entirely Color and location information in map interlinking models coupling cromogram is gathered to adjust the significance value of former Saliency maps picture using image Class and trellis search method search for the optimal weights of kernel function, and the Saliency maps picture in same cluster uses identical parameter setting It optimizes.This method is suitable for the optimization of a variety of conspicuousness detection algorithms, is applicable not only to Saliency maps under simple scenario Optimization is also applied for the optimization of Saliency maps under complex scene, and the Saliency maps after optimization more connect compared with former Saliency maps The result of person of modern times's work mark.
Description of the drawings
Fig. 1 is the flow diagram of the method for the present invention.
Fig. 2 is the implementation flow chart of one embodiment of the invention.
Specific implementation mode
Below in conjunction with the accompanying drawings and specific embodiment the invention will be further described.
The present invention provides a kind of conspicuousness inspection optimization method based on condition random field, as depicted in figs. 1 and 2, including Following steps：
Step S1：Extract the global depth convolution feature of each image in input picture set.Specifically include following steps：
Step S11：Extract image partial-depth convolution feature, this feature by image recognition depth network last layer Convolutional layer generates.In the present embodiment, the VGG- that image recognition depth network is won the championship using 2014 object identifications of ILSVRC match 19 depth networks.For the image I of input, the partial-depth convolution feature f of output d dimensions t × t.In the present embodiment, it is 512 The partial-depth convolution feature f of dimension 37 × 37.
Step S12：The aggregate weight of partial-depth convolution feature is calculated, calculation formula is as follows：
Wherein, α (x, y) indicates that the aggregate weight of pixel (x, y) in image I, σ take between picture centre and nearest boundary The 1/3 of distance.
Step S13：Weighting polymerization partial-depth convolution feature, obtains preliminary global characteristics, calculation formula is as follows：
Wherein,Indicate the preliminary global characteristics after image I polymerizations, dimension is 512；W and H indicates the width of image respectively Degree and height, in the present embodiment, W=H=37；F (x, y) indicates that the partial-depth convolution of pixel (x, y) in image I is special Sign.
Step S14：It is rightL2 standardization processings are carried out, calculation formula is as follows：
Wherein,It is the global characteristics after image I standardization, dimension is 512.
Step S15：It willPCA dimensionality reductions and whitening processing are carried out, the global depth convolution feature of n dimensions is obtainedThis N is 256 in embodiment, and calculation formula is as follows：
Wherein, MPCAIt is the PCA matrixes of a n × d, viIt is association singular value, i ∈ { 1,2 ..., n }, diag (v1, v2,...,vn) indicate diagonal matrix.
Step S16：It is rightL2 standardization processings are carried out, global depth convolution feature to the end is obtainedIt calculates public Formula is as follows：
Step S2：According to the similitude between image two-by-two in global depth convolution feature calculation input picture set.
The similitude between image is calculated using the dot product between corresponding global depth convolution feature two-by-two, calculation formula It is as follows：
Wherein,<>Indicate point multiplication operation, S (Ip,Iq) indicate image IpAnd IqBetween similitude, the value of similitude is bigger Indicate image IpAnd IqIt is more similar,Image I is indicated respectivelyp、IqGlobal depth convolution feature.
Step S3：K-means clusters are carried out to input picture set according to the similitude between image, form k mutually Independent image cluster.Specifically include following steps：
Step S31：The center that k images are image cluster is randomly selected, is then calculated by solving following optimization problem Other image IiAffiliated image cluster ci：
Wherein, UrIt is the center of image cluster, r=1,2 ..., k, S (Ii,Ur) it is image IiWith the center U of image clusterrBetween Similitude,Expression takes and image IiThe corresponding image cluster in center of the maximum image cluster of similitude, so After be assigned to ci, to image IiBelong to image cluster ci；
Step S32：The center of each image cluster takes the mean value of all images in the image cluster, forms new center, then It is clustered again by the step S31 methods for solving optimization problem；
Step S33：Continuous iterative step S32 obtains k image cluster { C until the center of each image cluster no longer changesr | r=1 ..., k }.
Step S4：The full condition of contact random field optimized parameter of each image cluster is calculated using trellis search method.Specifically Include the following steps：
Step S41：For every image in each image cluster, Saliency maps are generated using conspicuousness detection algorithm.
Step S42：For every image in each image cluster, it is random that full condition of contact is traversed using trellis search method The parameter ω of the energy function of field1、ω2、σα、σβ、σγAll valued combinations.The calculation formula of energy function is as follows：
Wherein, L indicates the notable label of whole image, L ∈ { 0,1 }W×H, { 0,1 }W×HIndicate W × H 0 or 1 value； E (L) indicates the corresponding energy functions of notable label L, liAnd ljThe notable label of pixel i and j in notable label L are indicated respectively, Value is that 0 expression is not notable, and value is that 1 expression is notable；P(li) be pixel i be notable label liProbability, take P (1)=δi, P (0)=1- δi, δiIt is the significance value of pixel i in Saliency maps；Energy function includes two parts：First partFor unit potential function, remaining part is pairs of potential function, including two cores, and first core is to be based on pixel The bilateral core of distance and color distortion, position is close and the similar similar significance value of pixel distribution of color, the core use Parameter ω1Control, pixel distance and color distortion use parameter σαAnd σβControl；Second core only depends on pixel distance, It aims at and improves isolated point small in former Saliency maps, using parameter ω2And σγControl；μ(li,lj) it is decision function, work as li =ljWhen, value 0, otherwise, value 1；piAnd pjThe pixel position of pixel i and j are indicated respectively；riAnd rjIt indicates respectively The color value of pixel i and j；The parameter ω1、ω2、σα、σβ、σγValue range be set of integers.In the present embodiment, The value range of parameter is：ω1The integer of value 3 to 6, ω2The integer of value 5 to 10, σαValue 3, σβBetween value 20 to 70 The integer that can be divided exactly by 10, σγThe integer and integer 33 that can be divided exactly by 5 between value 5 to 30, on the parameter traversals of grid search State all parameter values combination within the scope of parameter value.
For the parameter ω of every image1、ω2、σα、σβ、σγEach parameter value combination, complete calculate condition of contact with The minimization problem of the energy function on airport：
Expression takes the minimum corresponding notable label L of energy function E (L), is then assigned to significantly Label L *.
Each parameter value combination for every image, each picture in image is calculated according to the notable label L * solved The posterior probability of vegetarian refreshments, to generate the Saliency maps after optimization.
Step S43：Using accuracy rate recall rate area under the curve (Area Under Precision Recall Curve, PR-AUC) this basis for estimation of conspicuousness check and evaluation standard as optimized parameter valued combinations, for each image cluster, meter The average PR-AUC values for calculating Saliency maps after all image optimizations in the lower image cluster of each parameter value combination, take PR-AUC values Highest parameter value combines the full condition of contact random field optimized parameter as the image cluster.
Step S5：For new input picture, the image cluster belonging to it is judged, using the full condition of contact of affiliated image cluster Random field optimized parameter optimizes the Saliency maps of the new input picture.Specifically include following steps：
Step S51：New input picture I is generated using with identical conspicuousness detection algorithm in step S41eConspicuousness Figure；
Step S52：By solving following optimization problem calculating input image IeAffiliated image cluster ce：
Wherein, UrIt is the center for the image cluster that step S3 is obtained, S (Ie,Ur) it is input picture IeWith the center U of image clusterr Between similitude,Expression takes and input picture IeThe corresponding figure in center of the maximum image cluster of similitude As cluster, it is then assigned to ce, to input picture IeBelong to image cluster ce；
Step S53：Using image cluster ceFull condition of contact random field optimized parameter according to the calculation formula in step S42 To input picture IeSaliency maps optimize, generate optimization after Saliency maps.
The above are preferred embodiments of the present invention, all any changes made according to the technical solution of the present invention, and generated function is made When with range without departing from technical solution of the present invention, all belong to the scope of protection of the present invention.

Claims (6)

1. a kind of conspicuousness inspection optimization method based on condition random field, which is characterized in that include the following steps：
Step S1：Extract the global depth convolution feature of each image in input picture set；
Step S2：According to the similitude between image two-by-two in global depth convolution feature calculation input picture set；
Step S3：K-means clusters are carried out to input picture set according to the similitude between image, form k independently of each other Image cluster；
Step S4：The full condition of contact random field optimized parameter of each image cluster is calculated using trellis search method；
Step S5：For new input picture, the image cluster belonging to it is judged, the full condition of contact using affiliated image cluster is random Field optimized parameter optimizes the Saliency maps of the new input picture.
2. a kind of conspicuousness inspection optimization method based on condition random field according to claim 1, which is characterized in that institute It states in step S1, extracts the global depth convolution feature of image, include the following steps：
Step S11：Extract image partial-depth convolution feature, this feature by image recognition depth network last layer of convolution Layer generates, for the image I of input, the partial-depth convolution feature f of output d dimensions t × t；
Step S12：The aggregate weight of partial-depth convolution feature is calculated, calculation formula is as follows：
Wherein, α (x, y) indicates that the aggregate weight of pixel (x, y) in image I, σ take distance between picture centre and nearest boundary 1/3；
Step S13：Weighting polymerization partial-depth convolution feature, obtains preliminary global characteristics, calculation formula is as follows：
Wherein,Indicate the preliminary global characteristics after image I polymerizations, dimension is the width and height that d, W and H indicate image respectively Degree, f (x, y) indicate the partial-depth convolution feature of pixel (x, y) in image I；
Step S14：It is rightL2 standardization processings are carried out, calculation formula is as follows：
Wherein,Indicate the global characteristics after image I standardization, dimension is d；
Step S15：It willPCA dimensionality reductions and whitening processing are carried out, the global depth convolution feature of n dimensions is obtainedn<D is calculated Formula is as follows：
Wherein, MPCAIt is the PCA matrixes of a n × d, viIt is association singular value, i ∈ { 1,2 ..., n }, diag (v1,v2,..., vn) indicate diagonal matrix；
Step S16：It is rightL2 standardization processings are carried out, global depth convolution feature to the end is obtainedCalculation formula is as follows：
3. a kind of conspicuousness inspection optimization method based on condition random field according to claim 2, which is characterized in that institute It states in step S2, the similitude between image is calculated using the dot product between corresponding global depth convolution feature two-by-two, is calculated Formula is as follows：
Wherein,<>Indicate point multiplication operation, S (Ip,Iq) indicate image IpAnd IqBetween similitude, the bigger expression figure of the value of similitude As IpAnd IqIt is more similar,Image I is indicated respectivelyp、IqGlobal depth convolution feature.
4. a kind of conspicuousness inspection optimization method based on condition random field according to claim 3, which is characterized in that institute It states in step S3, K-means clusters is carried out to input picture set according to the similitude between image, are included the following steps：
Step S31：The center that k images are image cluster is randomly selected, it is then other by solving following optimization problem calculating Image IiAffiliated image cluster ci：
Wherein, UrIt is the center of image cluster, r=1,2 ..., k, S (Ii,Ur) it is image IiWith the center U of image clusterrBetween phase Like property,Expression takes and image IiThe corresponding image cluster in center of the maximum image cluster of similitude, then assignment To ci, to image IiBelong to image cluster ci；
Step S32：The center of each image cluster takes the mean value of all images in the image cluster, forms new center, then presses step The method that rapid S31 solves optimization problem is clustered again；
Step S33：Continuous iterative step S32 obtains k image cluster { C until the center of each image cluster no longer changesr| r= 1,...,k}。
5. a kind of conspicuousness inspection optimization method based on condition random field according to claim 4, which is characterized in that institute It states in step S4, the full condition of contact random field optimized parameter of each image cluster is calculated using trellis search method, including following Step：
Step S41：For every image in each image cluster, Saliency maps are generated using conspicuousness detection algorithm；
Step S42：For every image in each image cluster, full condition of contact random field is traversed using trellis search method The parameter ω of energy function1、ω2、σα、σβ、σγAll valued combinations；The calculation formula of energy function is as follows：
Wherein, L indicates the notable label of whole image, L ∈ { 0,1 }W×H, { 0,1 }W×HIndicate W × H 0 or 1 value；E(L) Indicate the corresponding energy functions of notable label L, liAnd ljThe notable label of pixel i and j in notable label L, value are indicated respectively Indicate not notable for 0, value is that 1 expression is notable；P(li) be pixel i be notable label liProbability, take P (1)=δi, P (0) =1- δi, δiIt is the significance value of pixel i in Saliency maps；Energy function includes two parts：First partFor unit potential function, remaining part is pairs of potential function, including two cores, and first core is to be based on pixel The bilateral core of distance and color distortion, position is close and the similar similar significance value of pixel distribution of color, the core use Parameter ω1Control, pixel distance and color distortion use parameter σαAnd σβControl；Second core only depends on pixel distance, It aims at and improves isolated point small in former Saliency maps, using parameter ω2And σγControl；μ(li,lj) it is decision function, work as li =ljWhen, value 0, otherwise, value 1；piAnd pjThe pixel position of pixel i and j are indicated respectively；riAnd rjIt indicates respectively The color value of pixel i and j；The parameter ω1、ω2、σα、σβ、σγValue range be set of integers；
For the parameter ω of every image1、ω2、σα、σβ、σγEach parameter value combination, calculate full condition of contact random field The minimization problem of energy function：
Expression takes the minimum corresponding notable label L of energy function E (L), is then assigned to notable label L *；
Each parameter value combination for every image, each pixel in image is calculated according to the notable label L * solved Posterior probability, to generate optimization after Saliency maps；
Step S43：Using accuracy rate recall rate area under the curve PR-AUC this conspicuousness check and evaluation standard as optimal ginseng The basis for estimation of number valued combinations calculates all images in the lower image cluster of each parameter value combination for each image cluster The average PR-AUC values of Saliency maps after optimization take the highest parameter value of PR-AUC values to combine the full connection as the image cluster Condition random field optimized parameter.
6. a kind of conspicuousness inspection optimization method based on condition random field according to claim 5, which is characterized in that institute It states in step S5, judges the image cluster belonging to new input picture, it is optimal using the full condition of contact random field of affiliated image cluster Parameter optimizes the Saliency maps of the new input picture, includes the following steps：
Step S51：New input picture I is generated using with identical conspicuousness detection algorithm in step S41eSaliency maps；
Step S52：By solving following optimization problem calculating input image IeAffiliated image cluster ce：
Wherein, UrIt is the center for the image cluster that step S3 is obtained, S (Ie,Ur) it is input picture IeWith the center U of image clusterrBetween Similitude,Expression takes and input picture IeThe corresponding image in center of the maximum image cluster of similitude Then cluster is assigned to ce, to input picture IeBelong to image cluster ce；
Step S53：Using image cluster ceFull condition of contact random field optimized parameter according to the calculation formula in step S42 to defeated Enter image IeSaliency maps optimize, generate optimization after Saliency maps.
CN201810256988.6A 2018-03-26 2018-03-26 Saliency detection optimization method based on conditional random field Active CN108491883B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810256988.6A CN108491883B (en) 2018-03-26 2018-03-26 Saliency detection optimization method based on conditional random field

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810256988.6A CN108491883B (en) 2018-03-26 2018-03-26 Saliency detection optimization method based on conditional random field

Publications (2)

Publication Number Publication Date
CN108491883A true CN108491883A (en) 2018-09-04
CN108491883B CN108491883B (en) 2022-03-22

Family

ID=63337515

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810256988.6A Active CN108491883B (en) 2018-03-26 2018-03-26 Saliency detection optimization method based on conditional random field

Country Status (1)

Country Link
CN (1) CN108491883B (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109740646A (en) * 2018-12-19 2019-05-10 创新奇智(北京)科技有限公司 A kind of image difference comparison method and its system, electronic device
CN109800692A (en) * 2019-01-07 2019-05-24 重庆邮电大学 A kind of vision SLAM winding detection method based on pre-training convolutional neural networks
CN111680702A (en) * 2020-05-28 2020-09-18 杭州电子科技大学 Method for realizing weak supervision image significance detection by using detection frame

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104103082A (en) * 2014-06-06 2014-10-15 华南理工大学 Image saliency detection method based on region description and priori knowledge
CN105574534A (en) * 2015-12-17 2016-05-11 西安电子科技大学 Significant object detection method based on sparse subspace clustering and low-order expression
CN107330431A (en) * 2017-06-30 2017-11-07 福州大学 A kind of conspicuousness inspection optimization method that fitting is clustered based on K means

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104103082A (en) * 2014-06-06 2014-10-15 华南理工大学 Image saliency detection method based on region description and priori knowledge
CN105574534A (en) * 2015-12-17 2016-05-11 西安电子科技大学 Significant object detection method based on sparse subspace clustering and low-order expression
CN107330431A (en) * 2017-06-30 2017-11-07 福州大学 A kind of conspicuousness inspection optimization method that fitting is clustered based on K means

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
ARTEM BABENKO ET AL: "Aggregating Local Deep Features for Image Retrieval", 《COMPUTER SCIENCE》 *
DONGJING SHAN ET AL: "Saliency optimization via low rank matrix recovery With Multi-Prior Integration", 《IEEE/CAA JOURNAL OF AUTOMATICA SINICA》 *
YUZHEN NIU ET AL: "CF‐based optimisation for saliency detection", 《IET COMPUTER VISION》 *
YUZHEN NIU ET AL: "Fitting-based optimisation for image visual salient object detection", 《IET COMPUTER VISION》 *
邓燕子等: "结合场景结构和条件随机场的道路检测", 《华中科技大学学报（自然科学版）》 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109740646A (en) * 2018-12-19 2019-05-10 创新奇智(北京)科技有限公司 A kind of image difference comparison method and its system, electronic device
CN109740646B (en) * 2018-12-19 2021-01-05 创新奇智(北京)科技有限公司 Image difference comparison method and system and electronic device
CN109800692A (en) * 2019-01-07 2019-05-24 重庆邮电大学 A kind of vision SLAM winding detection method based on pre-training convolutional neural networks
CN109800692B (en) * 2019-01-07 2022-12-27 重庆邮电大学 Visual SLAM loop detection method based on pre-training convolutional neural network
CN111680702A (en) * 2020-05-28 2020-09-18 杭州电子科技大学 Method for realizing weak supervision image significance detection by using detection frame
CN111680702B (en) * 2020-05-28 2022-04-01 杭州电子科技大学 Method for realizing weak supervision image significance detection by using detection frame

Also Published As

Publication number Publication date
CN108491883B (en) 2022-03-22

Similar Documents

Publication Publication Date Title
CN104408429B (en) A kind of video represents frame extracting method and device
WO2019192121A1 (en) Dual-channel neural network model training and human face comparison method, and terminal and medium
CN103530599B (en) The detection method and system of a kind of real human face and picture face
CN109903331B (en) Convolutional neural network target detection method based on RGB-D camera
CN108764041B (en) Face recognition method for lower shielding face image
CN107886507B (en) A kind of salient region detecting method based on image background and spatial position
CN108038476A (en) A kind of expression recognition feature extracting method based on edge detection and SIFT
CN108491883A (en) A kind of conspicuousness inspection optimization method based on condition random field
CN110276768B (en) Image segmentation method, image segmentation device, image segmentation apparatus, and medium
JP4098021B2 (en) Scene identification method, apparatus, and program
CN103984953A (en) Cityscape image semantic segmentation method based on multi-feature fusion and Boosting decision forest
WO2022077863A1 (en) Visual positioning method, and method for training related model, related apparatus, and device
CN107301643B (en) Well-marked target detection method based on robust rarefaction representation Yu Laplce&#39;s regular terms
CN109360179B (en) Image fusion method and device and readable storage medium
CN104346630A (en) Cloud flower identifying method based on heterogeneous feature fusion
Choi et al. Learning descriptor, confidence, and depth estimation in multi-view stereo
CN105469392B (en) High spectrum image conspicuousness detection method based on the comparison of region spectrum Gradient Features
CN110378893A (en) Image quality evaluating method, device and electronic equipment
CN107330431B (en) Significance detection optimization method based on K-means cluster fitting
CN111160107B (en) Dynamic region detection method based on feature matching
CN105389825B (en) Image processing method and system
CN111652260B (en) Face clustering sample number selection method and system
CN111178503A (en) Mobile terminal-oriented decentralized target detection model training method and system
CN106340008A (en) Feature value selection and SVM parameter optimization-based lung image recognition method
CN110910497B (en) Method and system for realizing augmented reality map

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