CN106650637A - Smiling face detector based on condition random forests and method - Google Patents

Smiling face detector based on condition random forests and method Download PDF

Info

Publication number
CN106650637A
CN106650637A CN201611096351.2A CN201611096351A CN106650637A CN 106650637 A CN106650637 A CN 106650637A CN 201611096351 A CN201611096351 A CN 201611096351A CN 106650637 A CN106650637 A CN 106650637A
Authority
CN
China
Prior art keywords
image
random forest
smiling face
face
head pose
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.)
Pending
Application number
CN201611096351.2A
Other languages
Chinese (zh)
Inventor
刘乐元
陈靓影
张坤
刘三女牙
杨宗凯
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.)
Huazhong Normal University
Original Assignee
Huazhong Normal 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 Huazhong Normal University filed Critical Huazhong Normal University
Priority to CN201611096351.2A priority Critical patent/CN106650637A/en
Publication of CN106650637A publication Critical patent/CN106650637A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168—Feature extraction; Face representation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/172—Classification, e.g. identification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/174—Facial expression recognition

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • General Health & Medical Sciences (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides a smiling face detector based on condition random forests. The smiling face detector comprises a condition random forest group, a face detection module, a head pose estimation module, a dynamic random forest construction module and a random forest smiling face classification module; the condition random forest group is obtained through offline training and composed of a group of smiling face classification random forests under different head pose conditions; the face detection module is used for detecting a face-containing region from an image; the head pose estimation module is used for estimating a head pose according to features extracted from the face region; the dynamic random forest construction module is used for selecting a decision-making tree from the condition random forest group to dynamically construct the random forests used for smiling face classification according to the estimated head pose; the random forest smiling face classification module is used for performing smiling face/non-smiling face classification on the image of the face region through the dynamically constructed smiling face classification random forests. Through the smiling face detector, smiling face detection under the situation that the head pose is not limited can be realized, and the application range of a smiling face detection system is expanded.

Description

A kind of smiling face's detector and method based on condition random forest
Technical field
The present invention relates to calculate machine vision and area of pattern recognition, and in particular to a kind of smiling face based on condition random forest examines Survey device and method.
Background technology
Smiling face's detection is in the side such as Consumer's Experience perception, students psychology state analysiss, photo enhancement process, camera smile shutter Face has a wide range of applications.Current smiling face's detection algorithm is mostly only effective to face image, and in many practical applications, people Head pose be generally free from limit.In the case where head pose excursion is larger, face alignment is more difficult, huge Difference can cause to be difficult to design high accuracy and efficient smiling face's detector in class.
The content of the invention
For the defect of prior art, laugh at it is an object of the present invention to solve to be detected in head pose image is not limited The problem of face, there is provided one kind possesses high accuracy, efficient smiling face's detector.
A kind of smiling face's detector based on condition random forest, including condition random forest group, face detection module, head Attitude estimation module, dynamic random forest set up module, random forest smiling face's sort module, wherein,
Condition random forest group is obtained by off-line training, random by the smiling face's classification under the conditions of one group of different head attitude Forest constitutes；Face detection module is used to detect the region comprising face from image；Head pose estimation module is used for root The feature assessment extracted according to human face region goes out head pose；Dynamic random forest sets up module to be used for according to the head pose estimated It is used for the random forest of smiling face's classification from trade-off decision tree dynamic construction in condition random forest group；Random forest smiling face classification mould Block is used to make human face region image using smiling face's classification random forest of dynamic establishment the classification of smiling face/non-smiling face.
Used as optimization, the condition random forest group off-line training is obtained, and the process of training is：
(1) image that training data is concentrated is divided into into N number of subset by the affiliated scope of head pose, is designated asWherein ΩnRepresent the head pose scope of the image included in n-th subset；
(2) using each data subsetIn image train a condition random forest T (Ωn), each condition random Forest T (Ωn) be made up of T decision tree, i.e.,Every decision tree Tt(Ωn) adopt following steps independence Training is generated：
(21) from training sample subsetMiddle random selection m image, then one is extracted from the every image chosen is Row characteristics of image, image IiExtract the characteristic image set after feature and be designated as Ii={ Ii 1,Ii 2..., Ii f..., Ii F, i ∈ [1, m], wherein F represent the kind number of feature；Then N is taken at random from every characteristic imagePIndividual size is the figure of s × s pixels As sub-block, and j-th image block taken from i-th image is designated as into Pij=(θi,Iij), wherein θi∈ { -1 ,+1 } is table Show the class label laughed at and ridicule；
(22) candidate's two-value test set is generatedEach two-value testComprising four parameters R1,R2, f, τ, its Middle R1And R2Two rectangular areas randomly selected from image subblock P are represented respectively, and f ∈ { 1,2 ..., F } are represented and randomly selected Feature passage, τ is threshold value；Each two-value testForm be：
Wherein IfRepresent and extract the image after f kind features, | R1|、|R2| pixel in two rectangular areas is represented respectively Quantity, (x, y) represents pixel；
(23) root node is generated, all image subblock { P for taking is included in the nodeij, root node is labeled as Present node；
(24) using each in candidate's two-value test setAttempt dividing the image subblock collection P on present node For two subsetsWith And
(25) information gain after division is calculated：
Wherein H () represents entropy；
When random tree grows into default depth capacity, or the information gain of present node be less than minimum threshold when, stop Only the growth of random tree and generate leaf node；Otherwise continue iteration division；The image subblock for reaching leaf node l is designated as l (P), count label in l (P) for smiling face (θ=1) Probability p (θ | l (P)), and p (θ | l (P)) is stored in into the leaf node On；
(3) each condition random forest classified threshold value is calculated：
(31) from training data subsetIn each image on take a series of images sub-block at random, and by each figure As sub-block input condition random forest T (Ωn) each decision tree, eventually arrive at the leaf node of decision tree；Take image subblock On the leaf node of arrival store class probability p (θ | Ωn, l (P)) and belong to the probability of smiling face as the image subblock.For table State and conveniently the probability that all image subblocks belong to smiling face is abbreviated as into { p1,p2,p3…}；
(32) cluster centre is initialized:c0=min { p1,p2,p3…},c1=max { p1,p2,p3…}；
(33) each data object p is calculatediTo cluster centre c0And c1Euclidean distance, and by each data be referred to away from From the class that shorter cluster centre is located, two class data acquisition systems after classification are designated as respectivelyWith
(34) all kinds of averages are calculated as new cluster centre；
(35) (32), (33) are repeated until cluster centre no longer changes；
(36) cluster result C is exported0、C1。
(37) design conditions random forest T (Ωn) classification thresholds τp(Ωn)：
Used as optimization, the dynamic random forest sets up condition random forest group of the module from off-line trainingIn Select T decision tree dynamic and set up out a random forest TC, wherein from condition random forest T (Ωn) in the decision-making selected at random A number k of treenThe head pose distribution situation be given by head pose estimation module is determining：
Wherein p (ω | P) the head pose distribution that head pose estimation module is given is represented,Expression is rounded downwards；It is dynamic The random forest T that state is set upCClassification thresholds τpFor：
Used as optimization, the random forest smiling face sort module receives the human face region figure from face detection module input Picture, and the random forest T set up using dynamicCThe classification of smiling face/non-smiling face is made to human face region image, step is：
1) to human face region image zooming-out feature；
2) M image subblock { P is taken respectively from all kinds of characteristic images of human face regioni, i ∈ [1, m] }, and be input into dynamic The random forest T that state is set upCEach decision-making seeds；When image subblock to PiReach decision treeLeaf node when, record Probit p preserved on the leaf nodeit=p (θ, l (P))；
If 3)The human face region image of judgement input is smiling face；Otherwise adjudicate the face of input Area image is non-smiling face.
A kind of smiling face detection method based on condition random forest, comprises the following steps：
Step 1) obtain being made up of the smiling face's classification random forest under the conditions of one group of different head attitude by off-line training Condition random forest group；
Step 2) region comprising face is detected from image；
Step 3) head pose is gone out according to the feature assessment of human face region extraction；
Step 4) smiling face is used for from trade-off decision tree dynamic construction in condition random forest group according to the head pose estimated The random forest of classification；
Step 5) smiling face's classification random forest for being set up using dynamic made smiling face/non-smiling face to human face region image and divides Class.
As optimization, the step 1) in condition random forest group off-line training obtain, the process of training is：
11) image that training data is concentrated is divided into into N number of subset by the affiliated scope of head pose, is designated asWherein ΩnRepresent the head pose scope of the image included in n-th subset；
12) using each data subsetIn image train a condition random forest T (Ωn), each condition random Forest T (Ωn) be made up of T decision tree, i.e.,Every decision tree Tt(Ωn) adopt following steps independence Training is generated：
121) from training sample subsetMiddle random selection m image, then one is extracted from the every image chosen is Row characteristics of image, image IiExtract the characteristic image set after feature and be designated as Ii={ Ii 1,Ii 2..., Ii f..., Ii F, i ∈ [1, m], wherein F represent the kind number of feature；Then N is taken at random from every characteristic imagePIndividual size is the figure of s × s pixels As sub-block, and j-th image block taken from i-th image is designated as into Pij=(θi,Iij), wherein θi∈ { -1 ,+1 } is table Show the class label laughed at and ridicule；
122) candidate's two-value test set is generatedEach two-value testComprising four parameters R1,R2, f, τ, its Middle R1And R2Two rectangular areas randomly selected from image subblock P are represented respectively, and f ∈ { 1,2 ..., F } are represented and randomly selected Feature passage, τ is threshold value；Each two-value testForm be：
Wherein IfRepresent and extract the image after f kind features, | R1|、|R2| pixel in two rectangular areas is represented respectively Quantity, (x, y) represents pixel；
123) root node is generated, all image subblock { P for taking is included in the nodeij, root node is labeled as Present node；
124) using each in candidate's two-value test setTrial is split into the image subblock collection P on present node Two subsetsWith And
125) information gain after division is calculated：
Wherein H () represents entropy；
When random tree grows into default depth capacity, or the information gain of present node be less than minimum threshold when, stop Only the growth of random tree and generate leaf node；Otherwise continue iteration division；The image subblock for reaching leaf node l is designated as l (P), count label in l (P) for smiling face (θ=1) Probability p (θ | l (P)), and p (θ | l (P)) is stored in into the leaf node On；
13) each condition random forest classified threshold value is calculated：
131) from training data subsetIn each image on take a series of images sub-block at random, and by each figure As sub-block input condition random forest T (Ωn) each decision tree, eventually arrive at the leaf node of decision tree；Take image subblock On the leaf node of arrival store class probability p (θ | Ωn, l (P)) and belong to the probability of smiling face as the image subblock.For table State and conveniently the probability that all image subblocks belong to smiling face is abbreviated as into { p1,p2,p3…}；
132) cluster centre is initialized:c0=min { p1,p2,p3…},c1=max { p1,p2,p3…}；
133) each data object p is calculatediTo cluster centre c0And c1Euclidean distance, and by each data be referred to away from From the class that shorter cluster centre is located, two class data acquisition systems after classification are designated as respectivelyWith
134) all kinds of averages are calculated as new cluster centre；
135) repeat 132), 133) until cluster centre no longer changes；
136) cluster result C is exported0、C1。
137) design conditions random forest T (Ωn) classification thresholds τp(Ωn)：
As optimization, the step 4) from the condition random forest group of off-line trainingIn select T decision tree Dynamic sets up out a random forest TC, wherein from condition random forest T (Ωn) in the number k of decision tree that selects at randomnBy The head pose distribution situation that head pose estimation module is given is determining：
Wherein p (ω | P) the head pose distribution that head pose estimation module is given is represented,Expression is rounded downwards；It is dynamic The random forest T that state is set upCClassification thresholds τpFor：
As optimization, the step 5) the human face region image being input into from face detection module is received, and utilize dynamic group The random forest T for buildingCThe classification of smiling face/non-smiling face is made to human face region image, step is：
51) to human face region image zooming-out feature；
52) M image subblock { P is taken respectively from all kinds of characteristic images of human face regioni, i ∈ [1, m] }, and be input into The random forest T that dynamic is set upCEach decision-making seeds；When image subblock to PiReach decision treeLeaf node when, note Record probit p preserved on the leaf nodeit=p (θ, l (P))；
If 53)The human face region image of judgement input is smiling face；The people for otherwise adjudicating input Face area image is non-smiling face.
As optimization, image subblock can not also be taken from whole human face region in training and detection process, and from mouth The regions such as bar, looks take image subblock, the technique construction smiling face's detector for then being provided using the present invention.
As optimization, also image subblock can be taken from multiple regions of multiple faces, the technology structure provided using the present invention The multiple detectors of smiling face are built, these smiling face's detectors are then combined, the accuracy of smiling face's detection is further improved.
The Advantageous Effects of the present invention are embodied in：
The invention provides a kind of smiling face's detecting system based on condition random forest.The system is being done using random forest Data are classified by head pose in training as cond according to head pose when smiling face detects, is reduced number According to interior class difference so that the condition random forest of training possesses high classification accuracy rate and efficiency；Carrying out smiling face/non-smiling face During classification, the estimation of head pose is also first carried out, random forest has been set up according to the head pose distribution situation dynamic estimated, and Calculate the classification thresholds of dynamic random forest.Therefore, the present invention provide technology can solution never limit under head pose scene Smiling face's test problems, extend smiling face's detecting system range of application.
Description of the drawings
Fig. 1 is a preferred embodiment of the present invention structure composition schematic diagram；
Fig. 2 is the off-line training flow chart of a preferred embodiment of the present invention condition random forest group；
Specific embodiment
In order that the objects, technical solutions and advantages of the present invention become more apparent, it is right below in conjunction with drawings and Examples The present invention is further elaborated.It should be appreciated that specific embodiment described herein is only to explain the present invention, and It is not used in the restriction present invention.As long as additionally, technical characteristic involved in invention described below each embodiment Not constituting conflict each other just can be mutually combined.
Fig. 1 shows a preferred embodiment of the present invention.A kind of smiling face's detector based on condition random forest, including： Condition random forest group 10, face detection module 11, head pose estimation module 12, dynamic random forest set up module 13, with Machine forest smiling face sort module 14, wherein,
Condition random forest group 10 is obtained by off-line training, by under the conditions of one group of different head attitude smiling face classification with Machine forest constitutes；Face detection module 11 detects the region comprising face from image；The basis of head pose estimation module 12 The feature assessment that human face region is extracted goes out head pose；Dynamic random forest sets up module 13 according to the head pose estimated from bar Trade-off decision tree dynamic construction in part random forest group is used for the random forest of smiling face's classification；Random forest smiling face sort module 14 The smiling face's classification random forest set up using dynamic makes the classification of smiling face/non-smiling face to human face region image.
The off-line training of condition random forest group 10 is obtained, and the process of training is as shown in Figure 2：
S11：The image that training data is concentrated is divided into into N number of subset by the affiliated scope of head pose, is designated asWherein ΩnRepresent the head pose scope of the image included in nth data subset.In the present embodiment In, training data is divided into into 3 subsets, i.e. Ω by horizontal deflection angle ω of head1={ -30 °≤ω≤30 ° }, Ω2= ω | -30 ° of ＜ ω≤- 60 ° } ∪ ω | 30 ° of ＜ ω≤60 ° } and Ω3=ω | ＜ -60 ° of -90 °≤ω } ∪ ω | 60 ° of ＜ ω≤90°}.Wherein, Ω2And Ω3The horizontal symmetry of face is make use of, the facial image for being oriented a left side is made after horizon glass picture Merge with the facial image for being oriented the right side.
S12：Using each data subsetIn image train a condition random forest T (Ωn).Each condition with Machine forest T (Ωn) be made up of T decision tree, i.e.,In the present embodiment, T is set to 30.Per certainly Plan tree Tt(Ωn) generated using following steps stand-alone training：
S121：From training sample subsetMiddle random selection m image.In the present embodiment, m is set to 2000.Then A series of images feature, image I are extracted from per an image choseniExtract the characteristic image set after feature and be designated as Ii= {Ii 1,Ii 2..., Ii F, i ∈ [1, m] wherein F represents the kind number of feature.In the present embodiment, F is set to 3, and employs 3 kinds Feature：Image intensity value, local binary pattern (LBP) and Gabor characteristic.Then N is taken at random from every characteristic imageP Individual size is the image subblock of s × s pixels, and j-th image block taken from i-th image is designated as into Pij=(θi,Iij), Wherein θi∈ { -1 ,+1 } (laughs at/ridicules) for class label.In the present embodiment, NPBe set to 200, s and be set to 25, i.e., from per The image subblock of 200 25 × 25 pixels is taken on characteristic image.
S122：Generate candidate's two-value test setEach two-value testComprising four parameters R1,R2, f, τ, Wherein R1And R2Two rectangular areas randomly selected from image subblock P are represented respectively, and f ∈ { 1,2 ..., F } represent random choosing The feature passage for taking, τ is threshold value.Each two-value testForm be：
Wherein | R1|、|R2| the quantity of pixel in rectangular area is represented respectively.In the present embodiment, candidate's two-value test set Be dimensioned to 3000.
S123：A root node is generated, all image subblock { P for taking are included in the nodeij}.Root node is labeled as Present node.
S124：Split vertexes.Using each in candidate's two-value test setAttempt image on present node Block collection P is split into two subsetsWith And
Calculate the information gain after division:
Wherein H () represents entropy.Select the two-value for causing information gain maximum test that present node is split into into two sons Node.
When random tree grows into default depth capacity, or the information gain of present node be less than minimum threshold when, stop Only the growth of random tree and generate leaf node；Otherwise continue iteration division.The image subblock for reaching leaf node l is designated as l (P), count label in l (P) for smiling face (θ=1) Probability p (θ | l (P)), and p (θ | l (P)) is stored in into the leaf node On.
S13：Calculate each condition random forest classified threshold value.The calculation procedure of threshold value is：
S131：From training data subsetIn each image on take a series of images sub-block at random, and by each figure As sub-block input condition random forest T (Ωn) each decision tree, eventually arrive at the leaf node of decision tree.Take image subblock On the leaf node of arrival store class probability p (θ | Ωn, l (P)) and belong to the probability of smiling face as the image subblock.For table State and conveniently the probability that all image subblocks belong to smiling face is abbreviated as into { p1,p2,p3…}。
S132：Initialization cluster centre:c0=min { p1,p2,p3…},c1=max { p1,p2,p3…}。
S133：Calculate each data object piTo cluster centre c0And c1Euclidean distance, and by each data be referred to away from From the class that shorter cluster centre is located.Two class data acquisition systems after classification are designated as respectivelyWith
S134：All kinds of averages are calculated as new cluster centre.
S135：Repeat (S132), (S133) until cluster centre no longer changes.
S136：Output cluster result C0、C1。
S137：Design conditions random forest T (Ωn) classification thresholds：
In the present embodiment, face detection module 11 is constituted by two based on the AdaBoost graders of Haar features, its In one be responsible for front face is detected from image, another be responsible for from image detect side face image.It can be appreciated that face Detection module 11 can also be implemented using other algorithms.
In the present embodiment, Algorithm of Head Pose Estimation adopts document《Robust head pose estimation using Dirichlet-tree distribution enhanced random forests》The algorithm of middle offer is realizing. The distribution situation of head pose is estimated by changing algorithm input human face region image.It can be appreciated that head pose detection module 12 can also can estimate algorithm that head pose is distributed implementing using other.
The dynamic random forest sets up condition random forest group 10 of the module 13 from off-line trainingIn select T Decision tree dynamic sets up out a random forest TC, wherein from condition random forest T (Ωn) in the decision tree selected at random Number knThe head pose distribution situation be given by head pose estimation module is determining：
Wherein p (ω | P) the head pose distribution that head pose estimation module is given is represented,Expression is rounded downwards.It is dynamic The random forest T that state is set upCClassification thresholds τpFor：
The random forest smiling face sort module 14 receives the human face region image from face detection module input, and utilizes The random forest T that dynamic is set upCThe classification of smiling face/non-smiling face is made to human face region image, step is：
(1) to human face region image zooming-out feature.The extraordinary species extracted and off-line training condition random forest group institute The feature species of extraction is consistent.
(2) M image subblock { P is taken respectively from all kinds of characteristic images of human face regioni, i ∈ [1, m] }, and be input into The random forest T that dynamic is set upCEach decision-making seeds.When image subblock to PiReach decision treeLeaf node when, note Record probit p preserved on the leaf nodeit=p (θ, l (P)).In the present embodiment M is set to 200.
(3) ifThe human face region image of judgement input is smiling face；Otherwise adjudicate the face of input Area image is non-smiling face.
As optimization, image subblock can not also be taken from whole human face region in training and detection process, and from mouth The regions such as bar, looks take image subblock, the technique construction smiling face's detector for then being provided using the present invention.As optimization, also Image subblock can be taken from multiple regions of multiple faces, the multiple detectors of technique construction smiling face provided using the present invention, so After combine these smiling face's detectors, further improve the accuracy of smiling face's detection.
This example can be implemented on the hardware such as including but not limited to smart mobile phone, panel computer, intelligent television, computer.
As it will be easily appreciated by one skilled in the art that the foregoing is only presently preferred embodiments of the present invention, not to The present invention, all any modification, equivalent and improvement made within the spirit and principles in the present invention etc. are limited, all should be included Within protection scope of the present invention.

Claims (8)

1. a kind of smiling face's detector based on condition random forest, it is characterised in that including condition random forest group, Face datection Module, head pose estimation module, dynamic random forest set up module, random forest smiling face's sort module, wherein,
Condition random forest group is obtained by off-line training, by the smiling face's classification random forest under the conditions of one group of different head attitude Composition；Face detection module is used to detect the region comprising face from image；Head pose estimation module is used for according to people The feature assessment of face extracted region goes out head pose；Dynamic random forest sets up module to be used for according to the head pose estimated from bar Trade-off decision tree dynamic construction in part random forest group is used for the random forest of smiling face's classification；Random forest smiling face sort module is used The classification of smiling face/non-smiling face is made to human face region image in the smiling face's classification random forest set up using dynamic.
2. the smiling face's detector based on condition random forest according to claim 1, it is characterised in that the condition random Forest group off-line training is obtained, and the process of training is：
(1) image that training data is concentrated is divided into into N number of subset by the affiliated scope of head pose, is designated as Wherein ΩnRepresent the head pose scope of the image included in n-th subset；
(2) using each data subsetIn image train a condition random forest T (Ωn), each condition random forest T(Ωn) be made up of T decision tree, i.e.,Every decision tree Tt(Ωn) adopt following steps stand-alone training Generate：
(21) from training sample subsetMiddle random selection m image, then extracts a series of figures from the every image chosen As feature, image IiExtract the characteristic image set after feature to be designated asWherein F represents the kind number of feature；Then N is taken at random from every characteristic imagePIndividual size is the image subblock of s × s pixels, and will J-th image block taken from i-th image is designated as Pij=(θi,Iij), wherein θi∈ { -1 ,+1 } is represented and is laughed at and ridicule Class label；
(22) candidate's two-value test set is generatedEach two-value testComprising four parameters R1,R2, f, τ, wherein R1 And R2Two rectangular areas randomly selected from image subblock P are represented respectively, and f ∈ { 1,2 ..., F } represent the spy for randomly selecting Passage is levied, τ is threshold value；Each two-value testForm be：
Wherein IfRepresent and extract the image after f kind features, | R1|、|R2| the quantity of pixel in two rectangular areas is represented respectively, (x, y) represents pixel；
(23) root node is generated, all image subblock { P for taking is included in the nodeij, root node is labeled as currently Node；
(24) using each in candidate's two-value test setImage subblock collection P on present node is split into two by trial SubsetWith And
(25) information gain after division is calculated：
Wherein H () represents entropy；
When random tree grows into default depth capacity, or the information gain of present node be less than minimum threshold when, stop with The growth of machine tree simultaneously generates leaf node；Otherwise continue iteration division；The image subblock for reaching leaf node l is designated as l (P), unites Label is the Probability p (θ | l (P)) of smiling face (θ=1) in meter l (P), and p (θ | l (P)) is stored on the leaf node；
(3) each condition random forest classified threshold value is calculated：
(31) from training data subsetIn each image on take a series of images sub-block at random, and by each image subblock Input condition random forest T (Ωn) each decision tree, eventually arrive at the leaf node of decision tree；Take image subblock arrival On leaf node store class probability p (θ | Ωn, l (P)) and belong to the probability of smiling face as the image subblock.It is convenient for statement The probability that all image subblocks belong to smiling face is abbreviated as into { p1,p2,p3…}；
(32) cluster centre is initialized:c0=min { p1,p2,p3…},c1=max { p1,p2,p3…}；
(33) each data object p is calculatediTo cluster centre c0And c1Euclidean distance, and by each data be referred to distance compared with The class that short cluster centre is located, two class data acquisition systems after classification are designated as respectivelyWith
(34) all kinds of averages are calculated as new cluster centre；
(35) (32), (33) are repeated until cluster centre no longer changes；
(36) cluster result C is exported0、C1。
(37) design conditions random forest T (Ωn) classification thresholds τp(Ωn)：
τ p ( Ω n ) = 1 2 ( m a x { p 1 0 , p 2 0 , p 3 0 , ... } + m i n { p 1 1 , p 2 1 , p 3 1 , ... } ) .
3. the smiling face's detector based on condition random forest according to claim 1, it is characterised in that the dynamic random Forest sets up condition random forest group of the module from off-line trainingIn select T decision tree dynamic and set up out one Random forest TC, wherein from condition random forest T (Ωn) in the number k of decision tree that selects at randomnBy head pose estimation mould The head pose distribution situation that block is given is determining：
Wherein p (ω | P) the head pose distribution that head pose estimation module is given is represented,Expression is rounded downwards；Dynamic is set up Random forest TCClassification thresholds τpFor：
τ p = τ p ( Ω n ) ∫ ω ∈ Ω n p ( ω | P ) d ω .
4. the smiling face's detector based on condition random forest according to claim 3, it is characterised in that the random forest Smiling face's sort module receives the human face region image from face detection module input, and the random forest T set up using dynamicCIt is right Human face region image makes the classification of smiling face/non-smiling face, and step is：
1) to human face region image zooming-out feature；
2) M image subblock { P is taken respectively from all kinds of characteristic images of human face regioni, i ∈ [1, m] }, and it is input into dynamic group The random forest T for buildingCEach decision-making seeds；When image subblock to PiReach decision treeLeaf node when, record the leaf Probit p preserved in child nodeit=p (θ, l (P))；
If 3)The human face region image of judgement input is smiling face；Otherwise adjudicate the human face region of input Image is non-smiling face.
5. a kind of smiling face detection method based on condition random forest, it is characterised in that comprise the following steps：
Step 1) obtain being classified the bar that constitutes of random forest by the smiling face under the conditions of one group of different head attitude by off-line training Part random forest group；
Step 2) region comprising face is detected from image；
Step 3) head pose is gone out according to the feature assessment of human face region extraction；
Step 4) classified for smiling face from trade-off decision tree dynamic construction in condition random forest group according to the head pose estimated Random forest；
Step 5) smiling face's classification random forest for being set up using dynamic makes the classification of smiling face/non-smiling face to human face region image.
6. the smiling face's detector based on condition random forest according to claim 5, it is characterised in that the condition random Forest group off-line training is obtained, and the process of training is：
11) image that training data is concentrated is divided into into N number of subset by the affiliated scope of head pose, is designated as Wherein ΩnRepresent the head pose scope of the image included in n-th subset；
12) using each data subsetIn image train a condition random forest T (Ωn), each condition random forest T(Ωn) be made up of T decision tree, i.e.,Every decision tree Tt(Ωn) adopt following steps stand-alone training Generate：
121) from training sample subsetMiddle random selection m image, then extracts a series of figures from the every image chosen As feature, image IiExtract the characteristic image set after feature to be designated asWherein F represents the kind number of feature；Then N is taken at random from every characteristic imagePIndividual size is the image subblock of s × s pixels, and will J-th image block taken from i-th image is designated as Pij=(θi,Iij), wherein θi∈ { -1 ,+1 } is represented and is laughed at and ridicule Class label；
122) candidate's two-value test set is generatedEach two-value testComprising four parameters R1,R2, f, τ, wherein R1 And R2Two rectangular areas randomly selected from image subblock P are represented respectively, and f ∈ { 1,2 ..., F } represent the spy for randomly selecting Passage is levied, τ is threshold value；Each two-value testForm be：
Wherein IfRepresent and extract the image after f kind features, | R1|、|R2| the quantity of pixel in two rectangular areas is represented respectively, (x, y) represents pixel；
123) root node is generated, all image subblock { P for taking is included in the nodeij, root node is labeled as currently Node；
124) using each in candidate's two-value test setImage subblock collection P on present node is split into two by trial Individual subsetWith And
125) information gain after division is calculated：
Wherein H () represents entropy；
When random tree grows into default depth capacity, or the information gain of present node be less than minimum threshold when, stop with The growth of machine tree simultaneously generates leaf node；Otherwise continue iteration division；The image subblock for reaching leaf node l is designated as l (P), unites Label is the Probability p (θ | l (P)) of smiling face (θ=1) in meter l (P), and p (θ | l (P)) is stored on the leaf node；
13) each condition random forest classified threshold value is calculated：
131) from training data subsetIn each image on take a series of images sub-block at random, and by each image subblock Input condition random forest T (Ωn) each decision tree, eventually arrive at the leaf node of decision tree；Take image subblock arrival On leaf node store class probability p (θ | Ωn, l (P)) and belong to the probability of smiling face as the image subblock.It is convenient for statement The probability that all image subblocks belong to smiling face is abbreviated as into { p1,p2,p3…}；
132) cluster centre is initialized:c0=min { p1,p2,p3…},c1=max { p1,p2,p3…}；
133) each data object p is calculatediTo cluster centre c0And c1Euclidean distance, and by each data be referred to distance compared with The class that short cluster centre is located, two class data acquisition systems after classification are designated as respectivelyWith
134) all kinds of averages are calculated as new cluster centre；
135) repeat 132), 133) until cluster centre no longer changes；
136) cluster result C is exported0、C1。
137) design conditions random forest T (Ωn) classification thresholds τp(Ωn)：
τ p ( Ω n ) = 1 2 ( m a x { p 1 0 , p 2 0 , p 3 0 , ... } + m i n { p 1 1 , p 2 1 , p 3 1 , ... } ) .
7. the smiling face detection method based on condition random forest according to claim 5, it is characterised in that the step 4) From the condition random forest group of off-line trainingIn select T decision tree dynamic and set up out a random forest TC, its In from condition random forest T (Ωn) in the number k of decision tree that selects at randomnThe head appearance be given by head pose estimation module State distribution situation is determining：
Wherein p (ω | P) the head pose distribution that head pose estimation module is given is represented,Expression is rounded downwards；Dynamic is set up Random forest TCClassification thresholds τpFor：
τ p = τ p ( Ω n ) ∫ ω ∈ Ω n p ( ω | P ) d ω .
8. the smiling face detection method based on condition random forest according to claim 5, it is characterised in that the step 5) Receive the human face region image from face detection module input, and the random forest T set up using dynamicCTo human face region image The classification of smiling face/non-smiling face is made, step is：
51) to human face region image zooming-out feature；
52) M image subblock { P is taken respectively from all kinds of characteristic images of human face regioni, i ∈ [1, m] }, and it is input into dynamic group The random forest T for buildingCEach decision-making seeds；When image subblock to PiReach decision tree Tt CLeaf node when, record the leaf Probit p preserved in child nodeit=p (θ, l (P))；
If 53)The human face region image of judgement input is smiling face；Otherwise adjudicate the face area of input Area image is non-smiling face.
CN201611096351.2A 2016-12-02 2016-12-02 Smiling face detector based on condition random forests and method Pending CN106650637A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201611096351.2A CN106650637A (en) 2016-12-02 2016-12-02 Smiling face detector based on condition random forests and method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201611096351.2A CN106650637A (en) 2016-12-02 2016-12-02 Smiling face detector based on condition random forests and method

Publications (1)

Publication Number Publication Date
CN106650637A true CN106650637A (en) 2017-05-10

Family

ID=58813864

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201611096351.2A Pending CN106650637A (en) 2016-12-02 2016-12-02 Smiling face detector based on condition random forests and method

Country Status (1)

Country Link
CN (1) CN106650637A (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108197225A (en) * 2017-12-28 2018-06-22 广东欧珀移动通信有限公司 Sorting technique, device, storage medium and the electronic equipment of image
CN108805009A (en) * 2018-04-20 2018-11-13 华中师范大学 Classroom learning state monitoring method based on multimodal information fusion and system
CN109598262A (en) * 2019-02-11 2019-04-09 华侨大学 A kind of children's facial expression recognizing method
CN109961148A (en) * 2019-04-03 2019-07-02 陕西医链区块链集团有限公司 Method for selecting block chain production nodes by using random forest algorithm
CN111523367A (en) * 2020-01-22 2020-08-11 湖北科技学院 Intelligent facial expression recognition method and system based on facial attribute analysis
CN111539333A (en) * 2020-04-24 2020-08-14 湖北亿咖通科技有限公司 Method for identifying gazing area and detecting distraction of driver

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140270551A1 (en) * 2013-03-15 2014-09-18 Nvidia Corporation Performing object detection operations via a graphics processing unit
KR101464446B1 (en) * 2013-10-07 2014-11-24 재단법인대구경북과학기술원 Method for user vertification using face recognition and head pose estimation and apparatus thereof
CN104504365A (en) * 2014-11-24 2015-04-08 闻泰通讯股份有限公司 System and method for smiling face recognition in video sequence
KR101506812B1 (en) * 2013-10-10 2015-04-08 재단법인대구경북과학기술원 Head pose estimation method using random forests and binary pattern run length matrix
CN104504366A (en) * 2014-11-24 2015-04-08 上海闻泰电子科技有限公司 System and method for smiling face recognition based on optical flow features
CN104680141A (en) * 2015-02-13 2015-06-03 华中师范大学 Motion unit layering-based facial expression recognition method and system
CN105787525A (en) * 2014-12-26 2016-07-20 小米科技有限责任公司 Image detection method and image detection device

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140270551A1 (en) * 2013-03-15 2014-09-18 Nvidia Corporation Performing object detection operations via a graphics processing unit
KR101464446B1 (en) * 2013-10-07 2014-11-24 재단법인대구경북과학기술원 Method for user vertification using face recognition and head pose estimation and apparatus thereof
KR101506812B1 (en) * 2013-10-10 2015-04-08 재단법인대구경북과학기술원 Head pose estimation method using random forests and binary pattern run length matrix
CN104504365A (en) * 2014-11-24 2015-04-08 闻泰通讯股份有限公司 System and method for smiling face recognition in video sequence
CN104504366A (en) * 2014-11-24 2015-04-08 上海闻泰电子科技有限公司 System and method for smiling face recognition based on optical flow features
CN105787525A (en) * 2014-12-26 2016-07-20 小米科技有限责任公司 Image detection method and image detection device
CN104680141A (en) * 2015-02-13 2015-06-03 华中师范大学 Motion unit layering-based facial expression recognition method and system

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
MATTHIAS DANTONE等: "Real-time facial feature detection using conditional regression forests", 《2012 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 *
YUANYUAN LIU等: "Robust head pose estimation using Dirichlet-tree distribution enhanced random forests", 《NEUROCOMPUTING》 *
ZHENZHEN LUO等: "Spontaneous Smile Recognition for Interest Detection", 《CHINESE CONFERENCE ON PATTERN RECOGNITION CCPR 2016: PATTERN RECOGNITION》 *
刘袁缘等: "基于树结构分层随机森林在非约束环境下的头部姿态估计", 《电子与信息学报》 *

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108197225A (en) * 2017-12-28 2018-06-22 广东欧珀移动通信有限公司 Sorting technique, device, storage medium and the electronic equipment of image
CN108197225B (en) * 2017-12-28 2020-12-01 Oppo广东移动通信有限公司 Image classification method and device, storage medium and electronic equipment
CN108805009A (en) * 2018-04-20 2018-11-13 华中师范大学 Classroom learning state monitoring method based on multimodal information fusion and system
CN109598262A (en) * 2019-02-11 2019-04-09 华侨大学 A kind of children's facial expression recognizing method
CN109961148A (en) * 2019-04-03 2019-07-02 陕西医链区块链集团有限公司 Method for selecting block chain production nodes by using random forest algorithm
CN109961148B (en) * 2019-04-03 2024-02-09 陕西医链区块链集团有限公司 Method for realizing selection of blockchain production nodes by using random forest algorithm
CN111523367A (en) * 2020-01-22 2020-08-11 湖北科技学院 Intelligent facial expression recognition method and system based on facial attribute analysis
CN111523367B (en) * 2020-01-22 2022-07-22 湖北科技学院 Intelligent facial expression recognition method and system based on facial attribute analysis
CN111539333A (en) * 2020-04-24 2020-08-14 湖北亿咖通科技有限公司 Method for identifying gazing area and detecting distraction of driver
CN111539333B (en) * 2020-04-24 2021-06-29 湖北亿咖通科技有限公司 Method for identifying gazing area and detecting distraction of driver

Similar Documents

Publication Publication Date Title
CN106650637A (en) Smiling face detector based on condition random forests and method
CN110443143B (en) Multi-branch convolutional neural network fused remote sensing image scene classification method
Angelova et al. Fast terrain classification using variable-length representation for autonomous navigation
CN111126360A (en) Cross-domain pedestrian re-identification method based on unsupervised combined multi-loss model
CN105930834B (en) Face identification method and device based on ball Hash binary-coding
CN107330390B (en) People counting method based on image analysis and deep learning
CN112906545B (en) Real-time action recognition method and system for multi-person scene
CN110008846B (en) Image processing method
CN111160286B (en) Video authenticity identification method
CN109376637A (en) Passenger number statistical system based on video monitoring image processing
CN106845513B (en) Manpower detector and method based on condition random forest
Wang et al. LiDAR data classification using morphological profiles and convolutional neural networks
Yogarajah et al. A dynamic threshold approach for skin tone detection in colour images
CN109635647B (en) Multi-picture multi-face clustering method based on constraint condition
CN106022223A (en) High-dimensional local-binary-pattern face identification algorithm and system
Samadzadegan et al. A multiple classifier system for classification of LIDAR remote sensing data using multi-class SVM
Chen et al. Cumulative attribute space regression for head pose estimation and color constancy
Timotius et al. Evaluation of edge orientation histograms in smile detection
CN104598898A (en) Aerially photographed image quick recognizing system and aerially photographed image quick recognizing method based on multi-task topology learning
CN104008374B (en) Miner's detection method based on condition random field in a kind of mine image
CN109886160A (en) It is a kind of it is non-limiting under the conditions of face identification method
Khryashchev et al. Audience analysis system on the basis of face detection, tracking and classification techniques
CN102156879B (en) Human target matching method based on weighted terrestrial motion distance
Sulistyaningrum et al. Vehicle detection using histogram of oriented gradients and real adaboost
CN113095235B (en) Image target detection method, system and device based on weak supervision and discrimination mechanism

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
RJ01 Rejection of invention patent application after publication

Application publication date: 20170510

RJ01 Rejection of invention patent application after publication