CN106156749A - Method for detecting human face based on selective search and device - Google Patents
Method for detecting human face based on selective search and device Download PDFInfo
- Publication number
- CN106156749A CN106156749A CN201610589408.6A CN201610589408A CN106156749A CN 106156749 A CN106156749 A CN 106156749A CN 201610589408 A CN201610589408 A CN 201610589408A CN 106156749 A CN106156749 A CN 106156749A
- Authority
- CN
- China
- Prior art keywords
- face
- region
- detected
- detection
- human face
- 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
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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30196—Human being; Person
- G06T2207/30201—Face
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 belongs to human face detection tech field, be specifically related to a kind of method for detecting human face based on selective search and device.The method for detecting human face based on selective search of the present invention comprises the following steps: extract face characteristic according to the positive and negative sample set of face, and training obtains face classification device；Based on selective search, image to be detected is carried out segmentation and obtain regional ensemble to be detected；Regional ensemble to be detected is expanded according to parameter preset；Face classification device is treated detection regional ensemble and is carried out detection and obtain human face region set；Filter out overlapping face and obtain final human face region set.Technical solution of the present invention by image to be detected is split, feature decision filter out the region being probably face, filter out meaningless region with this, significantly promote Face datection efficiency.
Description
Technical field
The invention belongs to human face detection tech field, be specifically related to a kind of method for detecting human face based on selective search and
Device.
Background technology
Face datection determines whether face exactly in a sub-picture or a sequence image (such as video), if having, returns
The information such as the size of the Huis' face, position.At present, human face detection tech has been widely used for various video monitoring place, wherein
Face classification device is most-often used technological means, and it refers to face and the non-face algorithm classified, by certain of image
Individual region input grader checks, it may be judged whether including face, different human face detection tech has different graders
Training method.
In actual application, likely occur in the optional position of image with arbitrary dimension due to face, thus necessary
Detecting each region of image, therefore, exhaustive search algorithm is typically to search for the common method of face location, and it is substantially
Flow process is as follows: first, inputs the positive and negative sample image of face, extracts face characteristic, and training obtains face classification device；Secondly, input
Image to be detected, begins stepping through entire image according to face classification device detection area size from image upper left position, by classification
Device testing result output human face region；Again, reduce image to be detected as input picture, repeat above-mentioned testing process, until
Input picture width higher primary school, in detection window, exports human face region set；Finally, overlapping face, output are filtered out with clustering algorithm
Final human face region.Said method shortcoming is: have detected all possible face location, face chi by traversal mode is exhaustive
Very little, treatment effeciency is low, so very time-consuming, especially on the limited embedded device of hardware resource, has algorithm performance
Tightened up requirement.
Summary of the invention
An object of the present invention is to overcome disadvantage mentioned above, it is achieved the optimization to Face datection step, improves face inspection
Survey efficiency, shorten the detection time.
In order to solve above-mentioned technical problem, the invention provides a kind of method for detecting human face based on selective search, bag
Include following steps:
Extracting face characteristic according to the positive and negative sample set of face, training obtains face classification device；
Based on selective search, image to be detected is carried out segmentation and obtain regional ensemble to be detected；
Regional ensemble to be detected is expanded according to parameter preset；
Face classification device is treated detection regional ensemble and is carried out detection and obtain human face region set；
Filter out overlapping face and obtain final human face region set.
Technical scheme, by selective search partitioning scheme, it is to avoid regions all to image travel through,
Drastically reduce the area the quantity of human face region to be detected, thus be effectively improved detection efficiency, shorten the detection time；Meanwhile, to dividing
Region after cutting carries out expanding the Face datection accuracy rate that can avoid causing because of segmentation result position, dimensional discrepancy and declines.
Further, described based on selective search, image to be detected is carried out segmentation obtain regional ensemble to be detected, bag
Include following steps:
Image to be detected is carried out primary segmentation；
Adjacent area is merged according to similarity condition；
The circumscribed rectangular region of combined region collection selected by frame；
Filter meaningless region；
The ratio of width to height revising region to be detected is consistent with face classification device detection zone territory the ratio of width to height.
By the region after primary segmentation being carried out according to similarity merging, and insignificant region is filtered,
All can reduce the final region quantity needing detection further, improve the efficiency of Face datection.
Further, described image to be detected is carried out primary segmentation before, further comprise the steps of: and image to be detected carried out
Reduce.
Technical scheme, by being reduced by image to be detected, on the premise of not affecting Detection results, carries
Rise splitting speed.
Further, described image to be detected is carried out primary segmentation, be to use pixel grey scale difference to cut rule.
Further, described similarity condition includes color similarity and size similarity.
Further, described parameter preset includes zooming parameter and offset parameter.
Further, described face classification device is treated detection regional ensemble and is carried out detection and obtain human face region set, including
Following steps:
It is high with face classification device detection zone field width consistent that area zoom to be detected is become；
Face classification device is treated detection region and is carried out Face datection, if detecting face, obtains human face region；
Repeat the above steps, until all region detection to be detected complete, obtains human face region set.
Further, described face characteristic is MB-LBP feature；Described face classification device is treated detection regional ensemble and is carried out
Detection obtains human face region set, particularly as follows:
Calculate image integration figure to be detected；
Face classification device detection region is scaled to high with region to be detected width consistent；
Calculate region to be detected all MB-LBP feature according to image integration figure to be detected, and input face classification device and carry out
Detection, if detecting face, obtains human face region；
Repeat the above steps, until all region detection to be detected complete, obtains human face region set.
When using MB-LBP feature as face characteristic, in technical scheme, use scaling face classification device
The method in detection region replaces scaling region to be detected, only need to calculate an integrogram, it is to avoid each scaling region to be detected
Time be required for recalculating the performance that integrogram causes and reduce；Meanwhile, on the equipment that support integrogram is hardware-accelerated, should
Scheme performance is more preferably.
Correspondingly, present invention also offers a kind of human face detection device based on selective search, including:
First processing module, for extracting face characteristic according to the positive and negative sample set of face, training obtains face classification device；
Second processing module, obtains set of regions to be detected for image to be detected being carried out segmentation based on selective search
Close；
3rd processing module, for expanding regional ensemble to be detected according to parameter preset；
Fourth processing module, treats detection regional ensemble for face classification device and carries out detection and obtain human face region set；
5th processing module, is used for filtering out overlapping face and obtains final human face region set.
Further, described second processing module, including:
First processing unit, for carrying out primary segmentation by image to be detected；
Second processing unit, for merging adjacent area according to similarity condition；
3rd processing unit, selects the circumscribed rectangular region of combined region collection for frame；
Fourth processing unit, is used for filtering meaningless region；
5th processing unit, for revising the ratio of width to height and the face classification device detection zone territory the ratio of width to height one in region to be detected
Cause.
Further, described fourth processing module, including:
First processing unit, high with face classification device detection zone field width consistent for area zoom to be detected is become；
Second processing unit, treating detection region for face classification device and carries out Face datection, if detecting face, obtaining
Human face region；
3rd processing unit, for repeat the above steps, until all region detection to be detected complete, obtains human face region
Set.
In sum, the beneficial effect of technical solution of the present invention has:
1. by selective search partitioning scheme, it is to avoid regions all to image travel through, and drastically reduce the area and treat
The quantity of detection human face region, thus it is effectively improved detection efficiency, shorten the detection time；By the region after primary segmentation is entered
Row merges according to similarity, and filters insignificant region, can reduce the final region needing detection further
Quantity, improves the efficiency of Face datection；Meanwhile, to segmentation after region carry out expansion can avoid because segmentation result position,
The Face datection accuracy rate that dimensional discrepancy causes declines.
2., by being reduced by image to be detected, on the premise of not affecting Detection results, promote splitting speed.
3., when using MB-LBP feature as face characteristic, use the method in scaling face classification device detection region to replace
Scale region to be detected, only need to calculate an integrogram, it is to avoid be required for recalculating one during each scaling region to be detected
The performance that secondary integrogram causes reduces；Meanwhile, on the equipment that support integrogram is hardware-accelerated, program performance is more preferably.
Accompanying drawing explanation
Fig. 1 is a kind of based on selective search the method for detecting human face flow chart of the present invention.
Fig. 2 is that the one of the present invention carries out segmentation based on selective search to image to be detected and obtains regional ensemble to be detected
Flow chart.
Fig. 3 is that a kind of face classification device of the present invention is treated detection regional ensemble and carried out detection and obtain human face region collective flow
Cheng Tu.
Fig. 4 is that the another kind of face classification device of the present invention is treated detection regional ensemble and carried out detection and obtain human face region set
Flow chart.
Fig. 5 is a kind of based on selective search the human face detection device structure chart of the present invention.
Fig. 6 is the second processing module structure chart of the present invention.
Fig. 7 is the fourth processing module structure chart of the present invention.
Detailed description of the invention
Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete
Describe, it is clear that described embodiment is only a part of embodiment of the present invention rather than whole embodiments wholely.Based on
Embodiment in the present invention, it is every other that those of ordinary skill in the art are obtained under not making creative work premise
Embodiment, broadly falls into the scope of protection of the invention.
Such as Fig. 1, it is a kind of based on selective search the method for detecting human face flow chart of the present invention, comprises the following steps:
Step 1, extracts face characteristic according to the positive and negative sample set of face, and training obtains face classification device；
Face datection determines whether face exactly in a sub-picture or a sequence image (such as video), if having, returns
The information such as the size of the Huis' face, position.In human face detection tech, face classification device is present stage most-often used technological means,
It refers to, by statistical method, the destination object of detection be carried out probability statistics, thus obtains some spies of object to be detected
Levying, it is established that target detection model, this modeling process is referred to as classifier training.After classifier training completes, just
Certain region of image can be inputted grader to detect, if be detected that include face, then the region of output matching,
Otherwise doing nothing, different human face detection tech has different classifier training methods.
The face classification device training method of technical solution of the present invention is: collects face training sample set and closes, training here
Sample is made up of according to certain ratio positive example sample and negative data, and wherein positive example sample refers to target sample to be checked, i.e. people
Face image, negative data refers to that any image of other non-face image, the width of all of sample image and height must be consistent；By institute
Some training samples input to grader, and grader extracts face characteristic from training sample, it is established that Face datection model, sample
This quantity is the most, and the face classification device Detection results that training obtains is the best.It addition, what the face classification device after having trained was supported
Detection zone field width height is to keep consistent with the wide high of input sample.
For example, it is possible to be the face sample set of 100 pixels by 300 wide height, (100 positive example samples and 200 are anti-
Example sample), training obtains detection zone field width 100 pixel, the face classification device of high 100 pixels.
Step 2, carries out segmentation based on selective search to image to be detected and obtains regional ensemble to be detected；
The optional position of image to be detected is likely occurred in due to face, so in traditional Face datection mode,
Must use exhaustive search algorithm that each region of image carries out traversal detection, to avoid detection region the most entirely to cause result not
Accurately；Simultaneously because face size is also unknown in image to be detected, there is the part that region to be detected is face
Situation, in this case, be necessary for image to be detected is constantly reduced, repeat above-mentioned testing process, until to be checked
Altimetric image narrows down in the same size with face classification device detection region, its is possible to detect complete face.This tradition people
Face detection mode is exhaustive have detected all possible face location, face size, and each positions and dimensions can once be divided
Class, so very time-consuming, especially on the limited embedded device of hardware resource, has tightened up wanting to algorithm performance
Ask.
Technical scheme reduces the region quantity needing detection to reach to improve by selective search mode
The purpose of detection performance, such as Fig. 2, is that the one of the present invention carries out segmentation based on selective search to image to be detected and obtains to be checked
Survey regional ensemble flow chart, comprise the following steps:
Step 201, carries out primary segmentation by image to be detected；
Image primary segmentation is exactly to become several to mutually disjoint image division to be detected according to certain rule, have solely
The region of property values, each region upper in some characteristic (such as gray scale, profile, color, texture etc.) and around region differ,
To analyze further and processing.Existing image partition method mainly divides following a few class: dividing method based on threshold value, based on
The dividing method in region, dividing method based on edge and dividing method based on particular theory etc..
In the preferred embodiments of the present invention, can use simple division rule based on pixel grey scale difference that image to be detected is entered
Row segmentation, the segmentation result obtained is more irregular area block.
Step 202, merges adjacent area according to similarity condition；
Image to be detected is carried out primary segmentation, is mainly brightness and the color of pixel in foundation image, but, due to figure
There is the reasons such as unsharp part and shade as in effect of noise, image, it occur frequently that segmentation errors, meanwhile, if
Threshold value arranges improper, also can cause the situation that cut zone block is too much.Based on above various reasons, at the beginning of image to be detected is carried out
On the basis of step segmentation, technical solution of the present invention also needs to adjacent area is carried out similarity judgement, if it is considered to this adjacent region
Territory is similar, then carry out region merging technique operation, and the process of merging is i.e. the process reducing set of regions to be detected.
In a specific embodiment, similarity condition can be color similarity, because relative to other feature, face
Color characteristic is highly stable, the most insensitive for rotation, translation and dimensional variation etc., and color characteristic calculates simple, so root
Carrying out combined region according to the color similarity degree of adjacent area is implementation more simply and easily.
In another specific embodiment, similarity condition can also be size similarity, according to the shape of adjacent area
And size similarity degree judges whether to need to merge.
Step 203, the circumscribed rectangular region of combined region collection selected by frame；
Owing to the region after merging still falls within irregular area block, and the detection region of face classification device needs are
Rectangular area, so the circumscribed rectangular region of combined region collection must also be selected as region to be detected by frame.The generally side of frame choosing
Method is the center reference location using the center of irregular area as boundary rectangle, and above-below direction and left and right directions select respectively
Select suitable length, it is ensured that all irregular areas are included in circumscribed rectangular region.
Step 204, filters meaningless region；
Owing to the wide high proportion of face generally remains in certain scope, and bigger with this proportional difference to be detected
Rectangular area is typically not facial image region, such as, the building wall in image, shade of object etc., this kind of we
Referred to as meaningless region, this part region to be detected can be filtered completely before testing, needs with further minimizing
Carry out the area results collection detected, improve the efficiency of detection.In the present invention, the wide high proportion parameter of face can configure,
Region to be detected beyond this ratio is carried out filtration treatment.
Such as, in the particular embodiment, the wide height of the face classification device detection window trained is 50 pixels,
The ratio of width to height is 1:1, and there are a width of 30 pixels in region to be detected, the rectangular area of a height of 150 pixels, and the ratio of width to height is 1:5,
The ratio of width to height in this region to be detected and the aspect ratio difference of normal face are relatively big, belong to meaningless region, can directly filter
Fall.
Step 205, the ratio of width to height revising region to be detected is consistent with face classification device detection zone territory the ratio of width to height.
Owing to the wide height in face classification device detection region is fixed value, when carrying out Face datection, also must in region to be detected
This size must be scaled to, input to the detection of face classification device.This just requires that all of region to be detected must be with face classification device
Detection region keep identical the ratio of width to height, arise that the situation that image stretch deform when of otherwise zooming in and out, affect
Detection accuracy.
The method that technical solution of the present invention uses is on the basis of regional center, and equal length is removed in left and right or up and down cutting
Region so that this region the ratio of width to height is consistent with grader the ratio of width to height.Such as, in a specific embodiment, face classification device is examined
Surveying a height of 50*60 of the wide * (unit picture element) of window, the ratio of width to height is 5:6, if one treats the modification region a height of 100*150 of width *, then repaiies
Positive way is: keeping wide 100 constant, on the basis of regional center, 15 pixels of cutting, obtain treating of 100*120 the most respectively
Detection region so that the ratio of width to height is 5:6；If another modification region a height of 140*120 of width *, then correcting mode is: keep high by 120
Constant, on the basis of regional center, left and right 20 pixels of cutting respectively, obtain the region to be detected of 100*120 so that the ratio of width to height
For 5:6.
In a preferred embodiment, image to be detected is split by technical solution of the present invention based on selective search
Obtain the step of regional ensemble to be detected, before image to be detected is carried out primary segmentation, it is also possible to include step: first will treat
Detection image reduces, by reducing input picture, to promote partitioning algorithm processing speed, such as, when to be detected
Time image is the big image in different resolution of 1080P, if directly artwork to be carried out partitioning algorithm process, when may expend more
Between, if artwork reduces 4 times, the widest height is all reduced into original 1/2nd, then partitioning algorithm processing speed will be greatly promoted,
Simultaneously because the wide high generally higher than 32*32 (unit picture element) of face classification device detection window, minimum face chi can be detected
Very little for 32*32 (unit picture element), after the wide height of artwork is all reduced into original 1/2nd, possible minimum face is a size of
16*16 (unit picture element), this area size is sufficiently large for partitioning algorithm, it is possible to correctly split, to follow-up
Face datection accuracy will not produce considerable influence.
Step 3, expands regional ensemble to be detected according to parameter preset；
Processing according to above-mentioned selective search mode after obtaining regional ensemble to be detected, technical solution of the present invention also needs to root
According to parameter preset, the set of regions to be detected being partitioned into is expanded, to improve the accuracy of Face datection.Such as, selectivity divides
Cut rear face and neck color, texture all very close to, it is easy to be assigned to same region, if face, neck will be comprised
Whole region input grader will be determined as non-face, just can arrange as long as this region at this time offsets up suitably distance
Remove the region of neck part, it may also be necessary to suitably scaling gets new detection region, the region input people new by this
Face grader just can detect that face.
Here parameter preset includes zooming parameter and offset parameter.Expanding set of regions to be detected according to zooming parameter is
Refer to, for a given region to be detected, treat detection region according to the ratio arranged and zoom in or out, again from treating
Detection image intercepts corresponding region and adds set of regions to be detected；Expand set of regions to be detected according to offset parameter to refer to, press
Arrange according to shift ratio laterally or longitudinally, from image to be detected, intercept corresponding region add set of regions to be detected.At this
In the specific embodiment of invention, zooming parameter and offset parameter can individually be arranged；Can also arrange simultaneously, zooming parameter and
Offset parameter comes into force simultaneously.
Such as, in a specific embodiment, zooming parameter is set as that { 0.8,1.1} times, offset parameter is set as { left and right
Direction: 0.1 times, above-below direction: 0.1 times }.For a given region to be detected (position in corresponding image to be detected is: in
Heart coordinate [120,100], region width is high [50,50]), include according to the set of regions that this region is expanded by zooming parameter:
Centre coordinate [120,100], region width is high [40,40]；(being reduced into 0.8 times)
Centre coordinate [120,100], region width is high [55,55]；(being enlarged into 1.1 times)
Include according to the set of regions that this region is expanded by offset parameter:
Centre coordinate [125,100], region width is high [50,50]；(left shift wide 0.1 times)
Centre coordinate [115,100], region width is high [50,50]；(offseting to the right wide 0.1 times)
Centre coordinate [120,105], region width is high [50,50]；(offsetting up high 0.1 times)
Centre coordinate [120,95], region width is high [50,50]；(offseting downward high 0.1 times)
It addition, the present embodiment is provided with zooming parameter and offset parameter simultaneously, it is necessary to carry out on the basis of scaling
Skew or zoom in and out on the basis of skew, all can get following extended area collection:
Centre coordinate [125,100], region width is high [40,40]；
Centre coordinate [115,100], region width is high [40,40]；
Centre coordinate [120,105], region width is high [40,40]；
Centre coordinate [120,95], region width is high [40,40]；
Centre coordinate [125,100], region width is high [55,55]；
Centre coordinate [115,100], region width is high [55,55]；
Centre coordinate [120,105], region width is high [55,55]；
Centre coordinate [120,95], region width is high [55,55].
The region of above-mentioned all expansions is added regional centralized to be detected, segmentation result position, dimensional discrepancy can be avoided
The accuracy rate caused declines, and improves the probability detecting face.
Step 4, face classification device is treated detection regional ensemble and is carried out detection and obtain human face region set；
Such as Fig. 3, it is that a kind of face classification device of the present invention is treated detection regional ensemble and carried out detection and obtain human face region collection
Close flow chart, comprise the following steps:
Step 401, becomes high with face classification device detection zone field width consistent by area zoom to be detected；
Due to set of regions to be detected be by several are the most consistent with the detection zone field width of face classification device but size can
Region that can be inconsistent forms, and the when that each region to be detected input face classification device detecting, needs district to be detected
The size scaling in territory becomes the consistent size in the detection region with face classification device, and whole region could be entered by such face classification device
The Face datection that row is complete.
Step 402, face classification device being treated detection region and carried out Face datection, if detecting face, obtaining human face region；
The detection model set up time face classification device is according to training, examines each region to be detected inputted
Surveying, if judging to meet face characteristic according to detection model, then obtaining the relevant letters such as face position within a detection region, size
Breath, and the human face region detected is exported.In actual application, from a region to be detected, can detect
Some multiple various sizes of human face regions.
Step 403, repeat the above steps, until all region detection to be detected complete, obtain human face region set.
Step 5, filters out overlapping face and obtains final human face region set.
The human face region set obtained after the detection of face classification device, is not the most final result, because these
Human face region there is likely to be the face of overlap.Such as, in aforesaid regional ensemble to be detected expands step, although pass through
Expand step and improve the accuracy of Face datection, too increase simultaneously and detect same face in multiple detection regions
Probability, therefore also needs to filter repeater's face, obtains final human face region set.
The method that present stage is commonly used is to filter out overlapping face, by face size, close the returning in position by clustering algorithm
It it is a face.In the particular embodiment, the classification of adjacent human face region can be carried out in the following way:
First according to formula: (adjacent human face region width minimum+adjacent human face region height is minimum for delta=0.25*
Value), calculate delta value, the most adjacent human face region width minimum, refer to width minimum in multiple adjacent human face region
Value, adjacent human face region height minima, refer to the value that multiple adjacent human face region camber is minimum.
If judging, between different human face regions, upper left corner horizontal stroke/vertical coordinate difference is less than delta and the human face region upper right corner again
Horizontal stroke/vertical coordinate difference is again smaller than delta, then these human face regions are classified as a class, belongs to same face, the face district after classification
Territory coordinate figure takes average.
Such as Fig. 4, it is that the another kind of face classification device of the present invention is treated detection regional ensemble and carried out detection and obtain human face region
Set flow chart, this technical scheme is the further optimization to above-mentioned method for detecting human face based on selective search, it is adaptable to
Using MB-LBP feature as the scene of face characteristic.Wherein, MB-LBP (Multi-Block Local BinaryPattern, base
In multilevel region local binary patterns) feature, refer to be divided into by facial image in several " sub-block regions ", each by thresholding
The average gray in " sub-block region " carries out feature extraction.
In using MB-LBP feature as the scene of face characteristic, face classification device is treated detection regional ensemble and is detected
Obtain human face region set, comprise the following steps:
Step 411, calculates image integration figure to be detected；
Integrogram is the most the most frequently used method quickly calculating MB-LBP eigenvalue, and its main thought is through time
Go through an image, image is arrived from the off each rectangular area pixel sum of being formed of point and saves as two-dimensional array, when
When calculating the MB-LBP feature in certain region, this area pixel and can obtain with the element of direct index array, need not weigh
New traversal calculate this region pixel and, improve calculating speed, MB-can be rapidly completed through four plus-minuss the most again
LBP feature calculation.
Step 412, is scaled to high with region to be detected width consistent by face classification device detection region；
In traditional Face datection mode using MB-LBP feature as face characteristic, owing to face size is to be detected
Image is unknown, it is necessary to image to be detected is constantly reduced, repeats detection.Contract at each image to be detected
After little, owing to image attributes there occurs change, the integrogram before calculated can not use, and needs again to travel through after calculating is reduced and treats
The integrogram of detection image, calculates MB-LBP feature laggard row face classification further according to new integrogram, repeatedly calculates integration
Figure can cause treatment effeciency low.
In technical scheme, the method for scaling face classification device detection window is used to go to mate region to be detected,
Keeping area size to be detected not change, there is not attribute change in region the most to be detected, still falls within image to be detected
A part, therefore in whole processing procedure, only need to calculate the image integration figure to be detected before once segmentation, follow-up each calculating
During the MB-LBP feature in region to be detected, may be by this integrogram, calculate image to be detected without again traveling through
Integrogram, substantially increases treatment effeciency.
Area size to be detected owing to obtaining after selective search processes not necessarily detects window with face classification device
Mouth is in the same size, but their wide high proportion is consistent, can be by the way of scaling face classification device detection window, it is achieved two
Person is in the same size, then carries out follow-up Face datection process.
Step 413, calculates region to be detected all MB-LBP feature according to image integration figure to be detected, and inputs face and divide
Class device detects, if detecting face, obtains human face region；
In actual face classification processes, may comprise many group MB-LBP features in a region, each MB-LBP is special
Levy corresponding to a rectangular area in region to be detected, may be overlapping between multiple rectangular areas, it is also possible to the most overlapping.Each
It is that the detection model generated by training determines that face classification device comprises how many group MB-LBP feature, such as, and the face trained
The information such as relative position within a detection region, each MB-LBP feature correspondence rectangular area, size are comprised inside detection model.
During face classification device detects, detection model can obtain the correspondence in region to be detected according to relative position, dimension information
Rectangular area, quickly calculates the MB-LBP feature of this rectangular area, carries out the differentiation of Face datection further according to integrogram, if inspection
Measure face, then obtain the relevant informations such as face position within a detection region, size, and the human face region detected is carried out
Output.
Step 414, repeat the above steps, until all region detection to be detected complete, obtain human face region set.
After obtaining human face region set, overlapping face can be filtered out by clustering algorithm equally and obtain final face district
Territory is gathered, and processes step with preceding solution consistent.
When using MB-LBP feature as face characteristic, technical scheme uses scaling face classification device detection
The method in region replaces scaling region to be detected, only need to calculate an integrogram, it is to avoid during each scaling region to be detected all
Need to recalculate the performance reduction that integrogram causes one time；Meanwhile, (such as sea on the equipment that support integrogram is hardware-accelerated
Think chip), after using integrogram hardware acceleration engine, this programme performance more preferably, can be rapidly completed the face to high-definition image
Detection.
Such as Fig. 5, it is a kind of based on selective search the human face detection device structure chart of the present invention, including:
First processing module, for extracting face characteristic according to the positive and negative sample set of face, training obtains face classification device；
By positive example sample and negative data are inputed to grader, grader extracts face characteristic from training sample,
Set up Face datection model.Face classification device after having trained can carry out detection to image and judge whether face.
Second processing module, obtains set of regions to be detected for image to be detected being carried out segmentation based on selective search
Close；The region quantity needing detection is reduced to reach to improve the purpose of detection performance by selective search mode.
3rd processing module, for expanding regional ensemble to be detected according to parameter preset；Can be according to zooming parameter and skew
Appointment region to be detected is expanded, to improve the accuracy of Face datection by parameter.
Fourth processing module, treats detection regional ensemble for face classification device and carries out detection and obtain human face region set；
Regional ensemble to be detected is inputted grader, and the detection model that grader uses training to obtain is treated detection region and is detected,
The human face region result set detected is exported.
5th processing module, is used for filtering out overlapping face and obtains final human face region set.
Such as Fig. 6, it is the second processing module structure chart of the present invention, including:
First processing unit, for carrying out primary segmentation by image to be detected；According to certain rule image to be detected
Be divided into several mutually disjoint, irregular area block with unique properties.
Second processing unit, for merging adjacent area according to similarity condition；According to color similarity, size similarity
Algorithm judges, if it is considered to adjacent area is similar, then carries out region merging technique operation, reduces set of regions to be detected.
3rd processing unit, selects the circumscribed rectangular region of combined region collection for frame；Irregular area after merging
Rectangular area elected as by frame.
Fourth processing unit, is used for filtering meaningless region；Owing to the wide high proportion of face generally remains in certain model
In enclosing, and the to be detected rectangular area bigger with this proportional difference is typically not facial image region, the most meaningless
Region, this part region to be detected can be filtered completely before testing, with the further district reducing and needing to carry out detecting
Field result collection, improves the efficiency of detection.
5th processing unit, for revising the ratio of width to height and the face classification device detection zone territory the ratio of width to height one in region to be detected
Cause, it is therefore an objective to time follow-up region to be detected or face classification device detection region zoom in and out, image stretch deformation will not be produced
Situation.
Such as Fig. 7, it is the fourth processing module structure chart of the present invention, including:
First processing unit, high with face classification device detection zone field width consistent for area zoom to be detected is become；
Second processing unit, treating detection region for face classification device and carries out Face datection, if detecting face, obtaining
Human face region；The detection model set up time face classification device is according to training, examines each region to be detected inputted
Surveying, if judging to meet face characteristic according to detection model, then obtaining the relevant letters such as face position within a detection region, size
Breath, and the human face region detected is exported.
3rd processing unit, for repeat the above steps, until all region detection to be detected complete, obtains human face region
Set.
Technical scheme is simply explained in detail by above-mentioned detailed description of the invention, the present invention the most only office
It is limited to above-described embodiment, every any improvement according to the principle of the invention or replacement, all should be within protection scope of the present invention.
Claims (11)
1. a method for detecting human face based on selective search, it is characterised in that comprise the following steps:
Extracting face characteristic according to the positive and negative sample set of face, training obtains face classification device；
Based on selective search, image to be detected is carried out segmentation and obtain regional ensemble to be detected；
Regional ensemble to be detected is expanded according to parameter preset；
Face classification device is treated detection regional ensemble and is carried out detection and obtain human face region set；
Filter out overlapping face and obtain final human face region set.
2. method for detecting human face based on selective search as claimed in claim 1, it is characterised in that described based on selectivity
Search carries out segmentation to image to be detected and obtains regional ensemble to be detected, comprises the following steps:
Image to be detected is carried out primary segmentation；
Adjacent area is merged according to similarity condition；
The circumscribed rectangular region of combined region collection selected by frame；
Filter meaningless region；
The ratio of width to height revising region to be detected is consistent with face classification device detection zone territory the ratio of width to height.
3. method for detecting human face based on selective search as claimed in claim 2, it is characterised in that described by mapping to be checked
Before carrying out primary segmentation, further comprise the steps of: and image to be detected is reduced.
4. method for detecting human face based on selective search as claimed in claim 2, it is characterised in that described by mapping to be checked
As carrying out primary segmentation, it is to use pixel grey scale difference to cut rule.
5. method for detecting human face based on selective search as claimed in claim 2, it is characterised in that described similarity condition
Including color similarity and size similarity.
6. method for detecting human face based on selective search as claimed in claim 1, it is characterised in that described parameter preset bag
Include zooming parameter and offset parameter.
7. method for detecting human face based on selective search as claimed in claim 1, it is characterised in that described face classification device
Treat detection regional ensemble to carry out detection and obtain human face region set, comprise the following steps:
It is high with face classification device detection zone field width consistent that area zoom to be detected is become；
Face classification device is treated detection region and is carried out Face datection, if detecting face, obtains human face region；
Repeat the above steps, until all region detection to be detected complete, obtains human face region set.
8. method for detecting human face based on selective search as claimed in claim 1, it is characterised in that described face characteristic is
MB-LBP feature；Described face classification device is treated detection regional ensemble and is carried out detection and obtain human face region set, particularly as follows:
Calculate image integration figure to be detected；
Face classification device detection region is scaled to high with region to be detected width consistent；
Calculate region to be detected all MB-LBP feature according to image integration figure to be detected, and input face classification device and examine
Surveying, if detecting face, obtaining human face region；
Repeat the above steps, until all region detection to be detected complete, obtains human face region set.
9. a human face detection device based on selective search, it is characterised in that including:
First processing module, for extracting face characteristic according to the positive and negative sample set of face, training obtains face classification device；
Second processing module, obtains regional ensemble to be detected for image to be detected being carried out segmentation based on selective search；
3rd processing module, for expanding regional ensemble to be detected according to parameter preset；
Fourth processing module, treats detection regional ensemble for face classification device and carries out detection and obtain human face region set；
5th processing module, is used for filtering out overlapping face and obtains final human face region set.
10. human face detection device based on selective search as claimed in claim 9, it is characterised in that described second processes
Module, including:
First processing unit, for carrying out primary segmentation by image to be detected；
Second processing unit, for merging adjacent area according to similarity condition；
3rd processing unit, selects the circumscribed rectangular region of combined region collection for frame；
Fourth processing unit, is used for filtering meaningless region；
5th processing unit, consistent with face classification device detection zone territory the ratio of width to height for revising the ratio of width to height in region to be detected.
11. human face detection device based on selective search as claimed in claim 9, it is characterised in that described fourth process
Module, including:
First processing unit, high with face classification device detection zone field width consistent for area zoom to be detected is become；
Second processing unit, treating detection region for face classification device and carries out Face datection, if detecting face, obtaining face
Region；
3rd processing unit, for repeat the above steps, until all region detection to be detected complete, obtains human face region collection
Close.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610589408.6A CN106156749A (en) | 2016-07-25 | 2016-07-25 | Method for detecting human face based on selective search and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610589408.6A CN106156749A (en) | 2016-07-25 | 2016-07-25 | Method for detecting human face based on selective search and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106156749A true CN106156749A (en) | 2016-11-23 |
Family
ID=58059790
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610589408.6A Pending CN106156749A (en) | 2016-07-25 | 2016-07-25 | Method for detecting human face based on selective search and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106156749A (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106937087A (en) * | 2017-02-07 | 2017-07-07 | 深圳云天励飞技术有限公司 | A kind of method for processing video frequency and device |
| CN107729889A (en) * | 2017-11-27 | 2018-02-23 | 广东欧珀移动通信有限公司 | Image processing method and device, electronic equipment, computer-readable recording medium |
| WO2018120460A1 (en) * | 2016-12-28 | 2018-07-05 | 平安科技（深圳）有限公司 | Image focal length detection method, apparatus and device, and computer-readable storage medium |
| CN108764352A (en) * | 2018-05-25 | 2018-11-06 | 百度在线网络技术（北京）有限公司 | Duplicate pages content detection algorithm and device |
| CN109214328A (en) * | 2018-08-29 | 2019-01-15 | 成都睿码科技有限责任公司 | Face grasping system based on face recognition engine |
| CN109389019A (en) * | 2017-08-14 | 2019-02-26 | 杭州海康威视数字技术股份有限公司 | Facial image selection method, device and computer equipment |
| CN109446984A (en) * | 2018-10-26 | 2019-03-08 | 重庆大学 | Traffic sign recognition method in natural scene |
| CN110619276A (en) * | 2019-08-15 | 2019-12-27 | 杭州电子科技大学 | Anomaly and violence detection system and method based on unmanned aerial vehicle mobile monitoring |
| CN111241891A (en) * | 2018-11-29 | 2020-06-05 | 中科视语(北京)科技有限公司 | Face image cutting method and device and computer readable storage medium |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20040037460A1 (en) * | 2002-08-22 | 2004-02-26 | Eastman Kodak Company | Method for detecting objects in digital images |
| CN101369315A (en) * | 2007-08-17 | 2009-02-18 | 上海银晨智能识别科技有限公司 | Human face detection method |
| CN103312963A (en) * | 2012-03-08 | 2013-09-18 | 索尼公司 | Image processing device and image processing method |
| CN103577838A (en) * | 2013-11-25 | 2014-02-12 | 苏州大学 | Face recognition method and device |
| CN103870834A (en) * | 2014-04-03 | 2014-06-18 | 张琰 | Method for searching for sliding window based on layered segmentation |
| CN104866844A (en) * | 2015-06-05 | 2015-08-26 | 中国人民解放军国防科学技术大学 | Crowd gathering detection method for monitor video |
| CN105335684A (en) * | 2014-06-25 | 2016-02-17 | 小米科技有限责任公司 | Face detection method and device |
-
2016
- 2016-07-25 CN CN201610589408.6A patent/CN106156749A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20040037460A1 (en) * | 2002-08-22 | 2004-02-26 | Eastman Kodak Company | Method for detecting objects in digital images |
| CN101369315A (en) * | 2007-08-17 | 2009-02-18 | 上海银晨智能识别科技有限公司 | Human face detection method |
| CN103312963A (en) * | 2012-03-08 | 2013-09-18 | 索尼公司 | Image processing device and image processing method |
| CN103577838A (en) * | 2013-11-25 | 2014-02-12 | 苏州大学 | Face recognition method and device |
| CN103870834A (en) * | 2014-04-03 | 2014-06-18 | 张琰 | Method for searching for sliding window based on layered segmentation |
| CN105335684A (en) * | 2014-06-25 | 2016-02-17 | 小米科技有限责任公司 | Face detection method and device |
| CN104866844A (en) * | 2015-06-05 | 2015-08-26 | 中国人民解放军国防科学技术大学 | Crowd gathering detection method for monitor video |
Non-Patent Citations (1)
| Title |
|---|
| 陈兆华: "基于视频的遮挡人脸检测与识别研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018120460A1 (en) * | 2016-12-28 | 2018-07-05 | 平安科技（深圳）有限公司 | Image focal length detection method, apparatus and device, and computer-readable storage medium |
| CN106937087A (en) * | 2017-02-07 | 2017-07-07 | 深圳云天励飞技术有限公司 | A kind of method for processing video frequency and device |
| CN109389019A (en) * | 2017-08-14 | 2019-02-26 | 杭州海康威视数字技术股份有限公司 | Facial image selection method, device and computer equipment |
| CN107729889B (en) * | 2017-11-27 | 2020-01-24 | Oppo广东移动通信有限公司 | Image processing method and device, electronic equipment and computer readable storage medium |
| CN107729889A (en) * | 2017-11-27 | 2018-02-23 | 广东欧珀移动通信有限公司 | Image processing method and device, electronic equipment, computer-readable recording medium |
| CN108764352A (en) * | 2018-05-25 | 2018-11-06 | 百度在线网络技术（北京）有限公司 | Duplicate pages content detection algorithm and device |
| CN108764352B (en) * | 2018-05-25 | 2022-09-27 | 百度在线网络技术（北京）有限公司 | Method and device for detecting repeated page content |
| CN109214328A (en) * | 2018-08-29 | 2019-01-15 | 成都睿码科技有限责任公司 | Face grasping system based on face recognition engine |
| CN109446984A (en) * | 2018-10-26 | 2019-03-08 | 重庆大学 | Traffic sign recognition method in natural scene |
| CN111241891A (en) * | 2018-11-29 | 2020-06-05 | 中科视语(北京)科技有限公司 | Face image cutting method and device and computer readable storage medium |
| CN111241891B (en) * | 2018-11-29 | 2024-04-30 | 中科视语(北京)科技有限公司 | Face image cutting method and device and computer readable storage medium |
| CN110619276B (en) * | 2019-08-15 | 2021-12-21 | 杭州电子科技大学 | Anomaly and violence detection system and method based on unmanned aerial vehicle mobile monitoring |
| CN110619276A (en) * | 2019-08-15 | 2019-12-27 | 杭州电子科技大学 | Anomaly and violence detection system and method based on unmanned aerial vehicle mobile monitoring |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106156749A (en) | Method for detecting human face based on selective search and device | |
| CN109961049B (en) | Cigarette brand identification method under complex scene | |
| CN108446617B (en) | Side face interference resistant rapid human face detection method | |
| WO2019140767A1 (en) | Recognition system for security check and control method thereof | |
| CN109635694B (en) | Pedestrian detection method, device and equipment and computer readable storage medium | |
| CN108898610A (en) | A kind of object contour extraction method based on mask-RCNN | |
| WO2018145470A1 (en) | Image detection method and device | |
| JP3073682B2 (en) | Method and apparatus for recognizing geometric shapes in an image | |
| CN109344702B (en) | Pedestrian detection method and device based on depth image and color image | |
| CN106709568A (en) | RGB-D image object detection and semantic segmentation method based on deep convolution network | |
| CN108549864A (en) | Area-of-interest filter method towards vehicle-mounted thermal imaging pedestrian detection and device | |
| CN103699900B (en) | Building horizontal vector profile automatic batch extracting method in satellite image | |
| CN105260749B (en) | Real-time target detection method based on direction gradient binary pattern and soft cascade SVM | |
| CN103473564B (en) | A kind of obverse face detection method based on sensitizing range | |
| CN104778481A (en) | Method and device for creating sample library for large-scale face mode analysis | |
| CN111967394B (en) | Forest fire smoke root node detection method based on dynamic and static grid fusion strategy | |
| CN102999886A (en) | Image edge detector and ruler raster grid line precision detection system | |
| CN104346801A (en) | Image-composition evaluating device, information processing device and method thereof | |
| CN102254159A (en) | Interpretation method for digital readout instrument | |
| CN103198479A (en) | SAR image segmentation method based on semantic information classification | |
| CN103955949A (en) | Moving target detection method based on Mean-shift algorithm | |
| CN108537816A (en) | A kind of obvious object dividing method connecting priori with background based on super-pixel | |
| CN106331746B (en) | Method and apparatus for identifying watermark location in video file | |
| CN105447834A (en) | Correction method for non-uniform illumination of mahjong images based on figure classification | |
| Su et al. | A new local-main-gradient-orientation HOG and contour differences based algorithm for object classification |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| CB02 | Change of applicant information | ||
| CB02 | Change of applicant information |
Address after: 350000, No. 2, building 9, 7 hi tech Avenue, Fuzhou Town, Minhou County, Fujian, China Applicant after: Fujian star net joint information system Co., Ltd. Address before: 350100, No. 2, building 9, 7 hi tech Avenue, Fuzhou Town, Minhou County, Fujian, China Applicant before: Fujian Star-Net Security Technology Co., Ltd. |
|
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20161123 |