CN108446602A - A kind of device and method for Face datection - Google Patents
A kind of device and method for Face datection Download PDFInfo
- Publication number
- CN108446602A CN108446602A CN201810166110.3A CN201810166110A CN108446602A CN 108446602 A CN108446602 A CN 108446602A CN 201810166110 A CN201810166110 A CN 201810166110A CN 108446602 A CN108446602 A CN 108446602A
- Authority
- CN
- China
- Prior art keywords
- face
- window
- layer model
- layer
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
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
- G06V40/172—Classification, e.g. identification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
- G06V40/161—Detection; Localisation; Normalisation
Abstract
The present invention provides a kind of device and method for Face datection.Described device includes：At least two-layer model, each layer model in addition to the first layer model are input with the output of its previous layer model；Wherein, first layer model is input with image to be detected, for screening the window that may include face from its input, and window of the possibility to being filtered out by it comprising face is calibrated, so that the rotation angle by the face in each window after the calibration is in the angular interval for the first layer model；Last layer model, for screening the window that may include face from its input, to export the result of Face datection.
Description
Technical field
The present invention relates to image procossings, more particularly to carry out Face datection to image.
Background technology
Whether Face datection refers to judging in given image including face, as a kind of image recognition technology
Known to more and more people.For example, when using mobile phone or digital camera shooting photo or video, middle appearance of finding a view is provided
Face position and size to carry out face U.S. face and auto-focusing, or for being in one section of video detection its picture
It is no face occur to be further processed, such as Face datection identity of personage, age, gender etc. based on appearance.People
The accuracy and detection speed of face detection have directly influenced the user experience of above application, however in many actual applied fields
Due to the influence of angle coverage and human body attitude in scape, face present in image to be detected is frequently not vertical, such as is deposited
Overhead downward, chin upward or there are the horizontal direction of face and image there are angles the case where, be Face datection reality
It applies and brings prodigious challenge.This requires the model of implementation Face datection or device while screening face that may be present,
Exclude the face that may be present compared between reference direction used in detection there are being influenced caused by angle, accurately
It detects the face existed in the image or judges in image whether to include face.
In view of the above-mentioned problems, some prior arts rotate the reference direction of detector to different directions, to avoid people
Influence of the rotation angle of face to Face datection.For example, being published in " IEEE Trans in 2007 in Chang Huang et al.
Article on Pattern Anal Mach Intell "《High-Performance Rotation Invariant
Multiview Face Detection》In, a unidirectional detector is had trained, then from four direction up and down
Four detectors are separately operable, to detect the face of arbitrary direction of rotation in plane.However, such Face datection mode
So that detection speed becomes very slow.
Also some prior arts propose to carry out Face datection to image to be detected using multi-stage cascade convolutional neural networks,
The calculating speed of convolutional neural networks is improved in such a way that multistage model is cascade, it is final to determine whether wrapped in image to be detected
Detection containing face.These in the prior art, in the predictive information of cascade network and not comprising rotation angle phase with face
The information of pass, although carrying out repeatedly iterative calculation by convolutional neural networks can reduce caused by the rotation angle of face
It influences, however this needs exchanges the standard of corresponding Face datection for using extremely large amount of iterative calculation and time cost as cost
True rate.
Also some prior arts, such as a kind of side of Face datection is proposed in patent document CN106529408A
Case is distributed a computing engines for each region to be scanned in image to be detected, and is matched in each engine
Dual-thread is set concurrently to be handled to improve the speed of Face datection.It is appreciated that for a width image to be detected, it can
It is uncertain that the face that can occur, which occupies the position of the ratio and face of image frame in the picture, thus is needed for institute
It states image to be detected and corresponding granularity is set to mark off a large amount of region to be scanned.For the above-mentioned prior art, due to
It distributes a computing engines for each region to be scanned, thus needs that a large amount of computing engines are arranged, and use is dedicated
FPGA hardware is realized.For common existing hardware device, it can not support the scheme of above-mentioned Face datection, need
Hardware is improved, and hardware cost will certainly be increased in this way.
In addition, there is some prior arts, such as used in patent document CN106778683A and CN107368797A
Tree-like human-face detector solves the problems, such as the Face datection of big angle rotary.For example, setting multilayer tree-shaped multi-orientation Face is examined
Device is surveyed, a grader is set in first layer to filter out the face being likely to occur；Two grader A are set in the second layer
And B, the reference direction of described two graders are arranged to opposite direction so that one of grader A is directed to rotation angle
It is detected for face upward, another grader B is that face directed downwardly is detected for rotation angle；In third layer
In, multiple graders are arranged in the output for grader A, with carry out in the second layer similarly further it is finer compared with
Multiple graders are arranged further to be divided, with such in the division of low-angle, the output for being similarly directed to grader B
It pushes away.However, in the above-mentioned methods in addition to first layer grader, the grader of remaining each layer is required to corresponding in preceding layer
As input, this makes in each layer in addition to the first layer for the output of grader, image to be detected it is same to be scanned
Region can repeatedly be handled in different graders.For example, the grader A in the second layer needs to handle first layer grader
The total data of output, and the grader B in the second layer is also required to the total data of processing first layer grader output.It can be with
Find out, such processing mode efficiency is very low, contains the calculating largely repeated.
Invention content
Therefore, it is an object of the invention to overcome the defect of the above-mentioned prior art, a kind of dress for Face datection is provided
It sets, including：
At least two-layer model, each layer model in addition to the first layer model are input with the output of its previous layer model；
Wherein, the first layer model is input with image to be detected, for screening the window that may include face from its input
Mouthful, and window of the possibility to being filtered out by it comprising face is calibrated, so that by each window after the calibration
Face rotation angle be in for the first layer model angular interval in；
Last layer model, for screening the window that may include face from its input, to export the knot of Face datection
Fruit.
Preferably, according to described device, wherein remaining in addition to first layer model and last described layer model
Each layer model includes face for screening the window that may include face from its input, and to the possibility filtered out by it
Window is calibrated, and current layer model is directed to so that being in by the rotation angle of the face in each window after the calibration
Angular interval in；
Wherein, the angular interval for current layer model is located within the angular interval for its previous layer model.
Preferably, according to described device, wherein carrying out calibration to the window that the possibility filtered out includes face includes：
According to may include face window in the rotation angle of face classify to the window；And
The rotation angle for being divided into face therein is deviated more from Face datection algorithm compared to other classifications to be adopted
The window of the classification of reference direction rotates corresponding angle；
Wherein, the angle rotated is arranged to opposite with the range of the rotation angle of face in the window of the classification
It answers.
Preferably, according to described device, wherein at least one layer in at least two-layer model uses convolutional neural networks
Model or SURF features and multilayer perceptron or HOG features and multilayer perceptron.
Preferably, according to described device, wherein each layer in at least two-layer model is arranged to mutual processing
Duration is same or similar.
Preferably, according to described device, further include：
At least one data buffer unit and control unit shared by adjacent two layers model；
The data buffer unit, for being written by one layer forward in the adjacent two layers model result being processed to
Wherein, and by one layer in the adjacent two layers model rearward data are therefrom read to be handled；
Described control unit, the result being processed in completion for controlling forward in the adjacent two layers model one layer
The data buffer unit is written and reads data corresponding with next image to be detected later.
And a method of Face datection is carried out using the device described in above-mentioned any one, including：
1) it may include face that first layer model carries out Face datection to filter out to image to be detected of input
Window, and window of the possibility to being filtered out by it comprising face is calibrated, so that passing through each window after the calibration
In face rotation angle be in for the first layer model angular interval in；
2) it may include face that last layer model carries out Face datection to screen to the content provided by previous layer model
Window, and export the result of Face datection.
Preferably, according to the method, wherein step 1) further includes：
Remaining each layer model in addition to first layer model and last described layer model by previous layer model to being carried
The content of confession carries out Face datection to screen the window that may include face, and includes the window of face to the possibility filtered out by it
Mouth is calibrated, so that the rotation angle by the face in each window after the calibration is in for current layer model
In angular interval；
Wherein, the angular interval for current layer model is located within the angular interval for previous layer model.
And a kind of computer readable storage medium, wherein being stored with computer program, the computer program is being held
For realizing above-mentioned any one method when row.
And a kind of system for Face datection, including：
Storage device and processor；
Wherein, for storing computer program, the computer program executes the storage device by the processor
When for realizing above-mentioned any one method.
Compared with the prior art, the advantages of the present invention are as follows：
Each layer model in human face detection device according to the present invention can be made whether the content of input to include people
The screening of face, this makes the quantity for the window being often retained after a layer model gradually decline, and there is no the same windows
The case where mouth is reprocessed by different graders, provides a kind of efficient Face datection scheme.
Also, human face detection device according to the present invention can gradually calibrate the rotation angle of each possible face,
The rotation angle of face will be carried out a degree of adjustment to realize that calibration, process are above-mentioned by often passing through the processing of a layer model
Further recognition of face is carried out to calibration rear hatch by next layer model again after adjustment.Face is gradually reduced as a result,
Maximum rotation angle is conducive to model and makes more accurate face and non-face screening.Design in this way, Face datection dress
It sets and can be used for carrying out for the face of arbitrary rotation angle in plane accurate and efficiently detect.
Human face detection device according to the present invention can both realize in a manner of hardware, can also be in the form of software
To realize.For by the way of realized using software, existing most processors can be compatible with, such as existing
Processing with Neural Network or general processor avoid increasing hardware cost.
In addition, human face detection device according to the present invention can also be in pipelined fashion continuously to multiple image (example
Such as video file) carry out Face datection.Each layer model is distributed to different computing units so that each computing unit is with flowing water
The mode of line is calculated, and to abundant utilization of hardware resources, further improves calculating speed.
Description of the drawings
Embodiments of the present invention is further illustrated referring to the drawings, wherein：
Fig. 1 shows the Face datection according to an embodiment of the invention for using three layers of Face datection and calibrating patterns
The structural schematic diagram of device；
Fig. 2 shows the flows of Face datection corresponding with the human face detection device in Fig. 1；
Fig. 3 shows the detection device using three layer model corresponding with Fig. 1 in pipelined fashion continuously to several
Image implements the assembly line sequence diagram of Face datection.
Specific implementation mode
It elaborates with reference to the accompanying drawings and detailed description to the present invention.
In the present invention, the cascade mode of multilayered model is equally used to form detection device, unlike the prior art
, each layer of the model in addition to last layer is required to detect whether there are possible face and according to may deposit
The rotation angle of face classify, last layer model can only detect whether that there are possible faces.To passing through one
The rotation angle of remaining region to be scanned (hereinafter referred to as " window ") carries out school generally after layer model screening
The result of calibration is input in next layer model by standard, with continue with the presence or absence of face detection and according to there may be
Face rotation angle classification.
Each layer model can reject the non-face window detected as a result, after the processing of each layer model
Remaining number of windows is successively decreased layer by layer.Also, calibration generally is carried out to the handling result of each layer model, enabling
The rotation angle for reducing possible face layer by layer reduces next layer and detects whether difficulty there are face, layer after being conducive to relatively
Model realize more accurately face and non-face differentiation to inputting window therein.It is to be understood that by leaning on relatively
The model of front layer carries out broad classification according to the rotation angle of face that may be present, by relatively rearward the model of layer according to more
Fine angle is added to classify, to when the result exported to each layer model executes calibration, realize to coarse, discrete
The prediction of direction and prediction to fine, continuous angle.
Human face detection device and its application method according to the present invention will be introduced by specific embodiment below.
Fig. 1 shows the Face datection according to an embodiment of the invention for using three layers of Face datection and calibrating patterns
The structural schematic diagram of device.It can be seen that three layers of Face datection and calibrating patterns are connected using cascade mode, equivalent layer
Model is inputted using the output of previous layer model as it.First layer Face datection and calibrating patterns are with the sliding window of image to be detected
Content (hereinafter referred to as " window ") in mouthful is inputted as it, and it is wherein non-face to exclude to carry out Face datection to each window
Window, and according to the rotation angle of face in each window carry out classification and according to the result of classification in each window
The rotation angle of face is substantially calibrated；Result after calibration is input into second layer Face datection and calibrating patterns, class
As, the second layer Face datection and calibrating patterns carry out Face datection to exclude wherein non-face window to each window,
And classified and calibrated according to the rotation angle of face in each window, and so on；Here second layer Face datection and
Calibrating patterns targeted rotation angle of classifying and calibrate is less than first layer Face datection and calibrating patterns, the inspection of third layer face
Survey and calibrating patterns, which classify and calibrate targeted rotation angle, is less than second layer Face datection and calibrating patterns, is achieved in
Gradual calibration to the rotation angle of face.
Due to, for most existing Face datection algorithms, when face rotation angle with reference to angle it
Between angle can realize when within 40 ° -60 ° and accurately identify very much, therefore multilayer people may be used in the present invention
Face detects and calibrating patterns are gradually to reduce face to be detected and carry out used by Face datection with reference between angle
Angle, with the accurate Face datection result of final output.
Fig. 2 shows the flows of Face datection corresponding with the human face detection device in Fig. 1.With reference to figure 2, using in Fig. 1
Human face detection device to image carry out Face datection method, including：
Whole windows of the first layer Face datection and calibrating patterns of human face detection device described in step 1. to image to be detected
Mouth carries out Face datection, and is divided into according to the rotation angle of the face for window of the possibility filtered out comprising face
Two classification, may be calibrated comprising the angular range of the window of face to this according to the classification divided.
If assuming that the rotation angle of face in image to be detected random distribution in all angles, can roughly by
The rotation angle of face be divided into " upward " and " downward " the two classify.If it will be understood, however, that being found by statistics
In the application scenarios of current face's detection, face appears in the rotation angle in image to be detected and Non-random distribution, then may be used also
Corresponding classification is arranged according to the result of statistics so that described two corresponding sections of classification and statistical result phase
Match.
As shown in Figure 2, by the screening of first layer Face datection and calibrating patterns, it may includes people to remain seven
The window of face.Here the face that the possibility that can be will identify that includes from its chin be directed toward the direction on its crown with straight up
Angle between reference direction (- 90 °, 90 °] in the range of window be divided into classification " upward ", the angle is existed
(90 °, 270 °] in the range of window be divided into classification " downward ".Thus, it is possible to being divided into classification " downward "
The rotation angle of face that may be present carries out rough calibration, such as is rotated 180 ° to be adjusted to classification " upward "
In.Such as in fig. 2, the possible face for including in first, fourth, five, six window be each divided into " downward " classification (with
Arrows go out the rotation angle of face), after their 180 ° of each spinnings, then these are originally belonged into classification " downward "
Windows calibration has arrived in classification " upward ".
In this way, the window that by the possibility detected can include face is all divided into the same classification
In, i.e., so that each window in possible face and reference direction between angle (- 90 °, 90 °] in the range of, and
And the reference direction of the same algorithm for being closer to Face datection of classifying, be conducive to pass through continuation in a subsequent step
Face datection is executed to obtain accurate recognition result.By mode illustrated in fig. 2, by the model of the rotation angle of script face
It encloses to have narrowed down to from 0 ° -360 ° and change in the range of 180 ° so that next layer model is in the rotation angle according to possible face
Degree can distinguish when being classified in the range of smaller.
The second layer Face datection and calibrating patterns of human face detection device described in step 2. by the first layer face to being examined
It surveys and the window of calibrating patterns output carries out Face datection, for window of the possibility further filtered out comprising face according to institute
The rotation angle for stating face is divided into two classification, and may include the angle of the window of face to this according to the classification divided
Range is further calibrated.
In step 2, the part input in second layer Face datection and calibrating patterns may have passed through the school of step 1
Standard has been adjusted in the section of smaller rotation angle, repeatedly implements people to window of this part after calibration at this time
Face detects, and can obtain higher accuracy rate.Here can face inspection only be carried out to the window for implementing calibration in step 1
It surveys, to improve computational efficiency.However, it is contemplated that the classification divided in step 1 according to rotation angle is very rough, therefore in reality
During the Face datection for applying second layer Face datection and calibrating patterns, preferably to by first layer Face datection and calibrating patterns
The window and be performed both by Face datection without the window of calibration that the process obtained after processing is calibrated.As shown in Figure 2, it is passing through
After having crossed the Face datection of second layer Face datection and calibrating patterns, six windows that may include face are remained.
With hereinbefore similarly, three classes can be divided into according to the range of the rotation angle of the face, if Face datection
The reference direction of algorithm be straight up, by rotation angle [- 90 °, -45 °) in the range of window be divided into " towards a left side "
Classification, window of the rotation angle in the range of [45 °, 90 °] is divided into classification " towards the right side ", by rotation angle [- 45 °,
45 °] in the range of window be divided into the classification of " temporarily without calibration ".
Here it is possible to the window for being divided into " towards a left side " is rotated clockwise 90 °, it is counterclockwise that window " towards the right side " will be divided into
It is rotated by 90 °, to which the rotation angle of possible face in whole windows to be calibrated in the range of [- 45 °, 45 °].
The third layer Face datection and calibrating patterns of human face detection device described in step 3. by the second layer face to being examined
It surveys and the window of calibrating patterns output carries out Face datection.Here third layer Face datection and calibrating patterns can be exported directly
Face datection as a result, for example identifying the window comprising face in the picture or the window comprising face being supplied to it
His software or hardware are further analyzed and handle.
If follow-up in current application also needs to face is identified or other processing, in step 3 can be with needle
Two classification are divided into according to the rotation angle of the face to the window that the possibility that further filters out includes face, and according to
The classification divided may further calibrate this comprising the angular range of the window of face, and export the knot after calibration
Fruit.With above-mentioned steps 1 and 2 similarly, can be by will be possible in whole windows according to the rotation angle of face in each window
The rotation angle of face is calibrated in the range of smaller.
The Face datection carried out to image to be detected may be implemented in 1-3 through the above steps as a result,.
Each layer model in above-described embodiment needs to realize that Face datection and the rotation angle based on face are classified,
Therefore present invention preferably employs convolutional neural networks to realize above-mentioned each layer of model, this is because convolutional neural networks are in reality
Extraordinary effect can be obtained when applying Face datection and classification.It will be understood, however, that there is some network models can also
Realization discriminates whether the effect that the rotation angle there are face and according to face is classified, such as SURF features (Speeded
Up Robust Features) and multilayer perceptron (Multi-layer Perceptron, MLP) or HOG features
(Histogram of Oriented Gradient) and multilayer perceptron.Therefore, in some embodiments of the invention, it is not necessary to
Each layer is all made of the model of same type, such as can be in first layer using SURF features and multilayer perceptron and the
Convolutional neural networks are used in two layers and third layer.
It will be understood to those skilled in the art that as described in the text, when the rotation angle of face and with reference between angle
Angle very accurately Face datection is advantageously implemented when within 40 ° -60 °.Therefore, face according to the present invention is being set
When the number of plies of the model of detection device, it is not necessary to three layer model is centainly used, as long as can be by the way that the rotation angle of face is gradually reduced
The mode of degree makes last layer when implementing Face datection, and the rotation angle of face is within 40 ° -60 ° in window.This
In in window the rotation angle of face the distribution of the rotation angle of face in application scenarios can be considered, such as pass through statistics
It was found that the rotation angle of 90% or more face is between -135 ° -135 ° in current application scene, then it can pass through setting
The model of the corresponding number of plies and the standard per layer model when implementing to classify gradually are calibrated the rotation angle of face with realizing
Within to ± 40 ° or within ± 60 °.
For example, according to other embodiments of the invention, the human face detection device with two-layer model can also be arranged, by
One layer model marks off four classification of " upward ", " to the left ", " to the right ", " downward " according to the rotation angle of face in window,
Correspond respectively to (- 45 °, 45 °], (45 °, 135 °], (135 °, 225 °], (225 °, -45 °] this four angular intervals；By second
Layer model realize in previous embodiment third layer Face datection and the consistent operation of calibrating patterns.
It is rough to realize by model layer forward in the above-mentioned human face detection device using multilayered model in view of it is expected
Face datection and broad classification, and it is expected to be identified accurately Face datection by model layer rearward and accurately classified
(calculation amount that classification needs higher fine degree and bigger is carried out in smaller range).It therefore, in the present invention can be with
The opposite model (such as model of first layer) for leaning on front layer is realized with the model of small-scale, is come with fairly large model real
Now with respect to the model of layer rearward (such as model of third layer).
Since most of the algorithm of Face datection is to improve accuracy rate by successive ignition, in other words, the rotation of face
Gyration is smaller, then the difficulty of Face datection is lower, it is thus possible to more rapidly reach desired accuracy rate.It can be appreciated that logical
Angle between the reference direction that face and algorithm use can gradually be reduced by crossing the above method of the present invention, be realized gradually
Promote calculating speed of each layer model to recognition of face.
Consider at above-mentioned 2 points, if the type and scale to each layer model select, it is likely that can make
The calculating time needed for per layer model is roughly the same, or at least on the same order of magnitude.For example, it is assumed that having three using above-mentioned
Detection device pair width image to be detected of layer network carries out Face datection, remaining window after being screened by each layer network
Quantity gradually decreases, such as by remaining 1000 windows after the first layer model, by being remained after the second layer model
100 windows can then make described two layers if the calculation amount of the second set layer model is 10 times of the first layer model
The calculating time of model is essentially identical.In the case, it is very beneficial for implementing each layer model by the way of assembly line
Operation, such as continuously implement Face datection for multiple image using each layer model as a thread of assembly line to improve
Method, this is because when being equal the processing time of each thread of assembly line assembly line efficiency highest.
In this regard, the present invention also provides a kind of using above-mentioned detection device in pipelined fashion continuously to multiple image
Implement the method for Face datection, such method is particularly suitable for being detected the face in video file.To video text
When part carries out Face datection, each layer model can be distributed to different computing units, calculated by the way of assembly line,
To abundant utilization of hardware resources, calculating speed is improved.
Fig. 3 shows the detection device using three layer model corresponding with Fig. 1 in pipelined fashion continuously to several
Image implements the assembly line sequence diagram of Face datection.With reference to figure 3, here respectively by first layer Face datection and calibrating patterns,
Two layers of Face datection and calibrating patterns and third layer Face datection and calibrating patterns, which are assigned to three different computing units, to be come in fact
It applies, is denoted as computing unit A, B and C respectively.In the present invention, used computing unit can be thread, CPU core, GPU cores
Etc. the unit for being commonly used for realization assembly line.
As shown in figure 3, in first unit interval T1, first layer Face datection and calibrating patterns reception first are to be checked
Altimetric image (input one A) and it is handled accordingly；
In second unit interval T2, first layer Face datection and second width image to be detected of calibrating patterns reception are (defeated
Enter two A) and it is handled accordingly, second layer Face datection and calibrating patterns are received from first layer Face datection
And calibrating patterns pending window (input one B) and it is handled accordingly；
In third unit interval T3, first layer Face datection and calibrating patterns reception third width image to be detected are (defeated
Enter three A) and it is handled accordingly, second layer Face datection and calibrating patterns are received from first layer Face datection
And calibrating patterns pending window (input two B) and it is handled accordingly, third layer Face datection and calibrating die
Type receives the pending window (three B of input) from second layer Face datection and calibrating patterns and is located accordingly to it
Reason；
And so on ...
As a result, by the way that the model of each layer to be distributed to different computing units, may be implemented continuously to multiple image into
Row processing, since each computing unit can work at the same time, it is thus possible to fully utilize the computing resource of hardware platform, significantly
Improve calculating speed.Also, a plurality of such assembly line can also be concurrently set in the present invention, calculated with further increasing
Speed.
The case where in view of for needing batch processing multiple image, handle each image needed for calculation amount there may be
Difference, it is likely that the computing unit that will appear next layer model not yet completes the processing to present image however previous layer model
Have been completed the case where processing to lower piece image.For example, computing unit A, B, C processing piece image is required to be more than
30ms, computing unit A, B, C handle the second width image and are required to about 10ms, then then there may be also exist as computing unit B
When handling piece image, computing unit A has been completed to the processing of the second width image and it is desirable that the result being processed to
It is supplied to computing unit B, then each level production line can be caused not to be connected normally in this way.
For such case, the present invention proposes data buffer zone can be arranged for each computing unit of assembly line, and
Corresponding control method is provided, to ensure the normal work of assembly line.Here data buffer zone is by two meters adjacent thereto
It is shared to calculate unit, such as the data buffer zone 1-2 between the first layer model and the second layer model, first layer mould
Data buffer zone 1-2 is written in the result that type can be processed to, and the second layer model can be needed from reading in the 1-2 of data buffer zone from it
The content of reason, write-in and reading here can be carried out at the same time so that when the first layer model handling result not by
All when write-in, the first layer model can be appreciated that the second layer model not yet completes the processing to preceding piece image, at this time first
Layer model will be postponed reading lower piece image.The above process can be realized by the control of system.
According to one embodiment of present invention, the workflow of the computing unit of each level production line includes：
The computing unit of the first order：First image to be detected for reading input, in the figure by way of sliding window
As upper some candidate regions of generation, it may includes face to carry out Face datection to the candidate region (i.e. window) to filter out
Window, and according to the rotation angle of setting to it is described may include face window in need the part calibrated to calibrate.
After completing the above process, in the shared data buffer zone of the computing unit of the computing unit and the second level of checking the first order
With the presence or absence of remaining space, if so, the data buffer zone is written into the handling result of the computing unit of the first order, if
It is no, then execute write operation after waiting for the data buffer zone remaining space occur.Until the first order computing unit it is complete
Portion's handling result has been written into after the data buffer zone, to be checked by next of the computing unit reading input of the first order
Altimetric image, and so on, to complete the processing to all image to be detected.
The second level to penultimate stage computing unit：From the data buffering shared by the computing unit of itself and previous stage
The content provided by the computing unit of previous stage is provided in area, is handled accordingly, that is, it may includes face to filter out
Window and the result of screening is calibrated.Similarly with the computing unit of the above-mentioned first order, after completing the above process,
Write-in warp is chosen whether according to whether having remaining space in the data buffer zone shared by the computing unit of itself and rear stage
Its treated result.After completing above-mentioned write operation, delay from the data shared by the computing unit of itself and previous stage
Reading of content in area is rushed, is handled, the above process is repeated, until there is no data transfers to come.
The computing unit of afterbody：Analogously with above-mentioned computing unit, corresponding processing operation is carried out, is directed to
The Face datection of current image to be detected as a result, directly exporting the result.Then, its computing unit with previous stage is read
Content in the data buffer zone shared is handled, and the above process is repeated, until there is no data transfers to come.
In this way, each computing unit that can control assembly line is protected on reading data and processing data
Hold synchronization, avoid the loss of data, ensure assembly line can normally work.
In conclusion the present invention provides a kind of improved Face datection scheme, it can be directed in plane and arbitrarily rotate
The face of angle is carried out accurate and is efficiently detected.Human face detection device according to the present invention can be with the side of hardware or software
Formula is realized, existing most processors can be compatible with when realizing in the form of software, avoid increasing hardware cost.And
And human face detection device according to the present invention can also be in pipelined fashion continuously to the multiple image of such as video file
Face datection is carried out, abundant utilization of hardware resources further improves calculating speed.
It should be noted that each step introduced in above-described embodiment is all not necessary, those skilled in the art
Can carry out according to actual needs it is appropriate accept or reject, replace, modification etc..
It should be noted last that the above examples are only used to illustrate the technical scheme of the present invention and are not limiting.On although
Text is described the invention in detail with reference to embodiment, it will be understood by those of ordinary skill in the art that, to the skill of the present invention
Art scheme is modified or replaced equivalently, and without departure from the spirit and scope of technical solution of the present invention, should all be covered at this
In the right of invention.
Claims (10)
1. a kind of device for Face datection, including：
At least two-layer model, each layer model in addition to the first layer model are input with the output of its previous layer model；
Wherein, the first layer model is input with image to be detected, for screening the window that may include face from its input, and
The window that the possibility filtered out by it includes face is calibrated, so that passing through the face in each window after the calibration
Rotation angle be in for the first layer model angular interval in；
Last layer model, for screening the window that may include face from its input, to export the result of Face datection.
2. the apparatus according to claim 1, wherein its in addition to first layer model and last described layer model
Remaining each layer model includes face for screening the window that may include face from its input, and to the possibility filtered out by it
Window calibrated be directed to current layer mould so that being in by the rotation angle of the face in each window after the calibration
In the angular interval of type；
Wherein, the angular interval for current layer model is located within the angular interval for its previous layer model.
3. the apparatus according to claim 1, wherein carrying out calibration to the window that the possibility filtered out includes face includes：
According to may include face window in the rotation angle of face classify to the window；And
Used by the rotation angle for being divided into face therein is deviated more from Face datection algorithm compared to other classifications
The window of the classification of reference direction rotates corresponding angle；
Wherein, the angle rotated is arranged to corresponding with the range of the rotation angle of face in the window of the classification.
4. according to the device described in any one of claim 1-3, wherein at least one layer in at least two-layer model is adopted
With convolutional neural networks model or SURF features and multilayer perceptron or HOG features and multilayer perceptron.
5. according to the device described in any one of claim 1-3, wherein each layer in at least two-layer model is set
It is same or similar to be set to mutual handling duration.
6. according to the device described in any one of claim 1-3, further include：
At least one data buffer unit and control unit shared by adjacent two layers model；
The data buffer unit, for it to be written by one layer of result being processed to forward in the adjacent two layers model
In, and data are therefrom read by one layer in the adjacent two layers model rearward and are handled；
Described control unit, for controlling forward in the adjacent two layers model one layer in the result write-in for completing to be processed to
Data corresponding with next image to be detected are read after the data buffer unit.
7. a kind of method that device using described in any one of claim 1-6 carries out Face datection, including：
1) first layer model carries out Face datection to filter out the window that may include face to image to be detected of input,
And window of the possibility to being filtered out by it comprising face is calibrated, so that passing through the people in each window after the calibration
The rotation angle of face is in the angular interval for the first layer model；
2) last layer model carries out Face datection to screen the window that may include face to the content provided by previous layer model
Mouthful, and export the result of Face datection.
8. according to the method described in claim 7, wherein step 1) further includes：
Remaining each layer model in addition to first layer model and last described layer model by previous layer model to being provided
Content carries out Face datection to screen the window that may include face, and window of the possibility comprising face to being filtered out by it into
Row calibration, so that being in the angle for current layer model by the rotation angle of the face in each window after the calibration
In section；
Wherein, the angular interval for current layer model is located within the angular interval for previous layer model.
9. a kind of computer readable storage medium, wherein being stored with computer program, the computer program is used when executed
In realization method as claimed in claim 7 or 8.
10. a kind of system for Face datection, including：
Storage device and processor；
Wherein, the storage device for storing computer program, when being executed by the processor use by the computer program
In realization method as claimed in claim 7 or 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810166110.3A CN108446602B (en) | 2018-02-28 | 2018-02-28 | Device and method for detecting human face |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810166110.3A CN108446602B (en) | 2018-02-28 | 2018-02-28 | Device and method for detecting human face |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108446602A true CN108446602A (en) | 2018-08-24 |
| CN108446602B CN108446602B (en) | 2021-08-20 |
Family
ID=63192710
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810166110.3A Active CN108446602B (en) | 2018-02-28 | 2018-02-28 | Device and method for detecting human face |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108446602B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109635755A (en) * | 2018-12-17 | 2019-04-16 | 苏州市科远软件技术开发有限公司 | Face extraction method, apparatus and storage medium |
| WO2021174688A1 (en) * | 2020-03-05 | 2021-09-10 | 平安科技（深圳）有限公司 | Facial detection method and system |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2001043349A (en) * | 1999-07-27 | 2001-02-16 | Fujitsu Ltd | Face posture detector |
| CN101639933A (en) * | 2009-07-16 | 2010-02-03 | 上海合合信息科技发展有限公司 | Image rotation correction method and system and electronic device |
| CN106485215A (en) * | 2016-09-29 | 2017-03-08 | 西交利物浦大学 | Face occlusion detection method based on depth convolutional neural networks |
| CN107368797A (en) * | 2017-07-06 | 2017-11-21 | 湖南中云飞华信息技术有限公司 | The parallel method for detecting human face of multi-angle, device and terminal device |
| CN107506707A (en) * | 2016-11-30 | 2017-12-22 | 奥瞳系统科技有限公司 | Using the Face datection of the small-scale convolutional neural networks module in embedded system |
-
2018
- 2018-02-28 CN CN201810166110.3A patent/CN108446602B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2001043349A (en) * | 1999-07-27 | 2001-02-16 | Fujitsu Ltd | Face posture detector |
| CN101639933A (en) * | 2009-07-16 | 2010-02-03 | 上海合合信息科技发展有限公司 | Image rotation correction method and system and electronic device |
| CN106485215A (en) * | 2016-09-29 | 2017-03-08 | 西交利物浦大学 | Face occlusion detection method based on depth convolutional neural networks |
| CN107506707A (en) * | 2016-11-30 | 2017-12-22 | 奥瞳系统科技有限公司 | Using the Face datection of the small-scale convolutional neural networks module in embedded system |
| CN107368797A (en) * | 2017-07-06 | 2017-11-21 | 湖南中云飞华信息技术有限公司 | The parallel method for detecting human face of multi-angle, device and terminal device |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109635755A (en) * | 2018-12-17 | 2019-04-16 | 苏州市科远软件技术开发有限公司 | Face extraction method, apparatus and storage medium |
| WO2021174688A1 (en) * | 2020-03-05 | 2021-09-10 | 平安科技（深圳）有限公司 | Facial detection method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108446602B (en) | 2021-08-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Li et al. | Instance-level salient object segmentation | |
| KR102236046B1 (en) | Face detection training method, device and electronic device | |
| WO2020164282A1 (en) | Yolo-based image target recognition method and apparatus, electronic device, and storage medium | |
| CN108470354A (en) | Video target tracking method, device and realization device | |
| CN110321873B (en) | Sensitive picture identification method and system based on deep learning convolutional neural network | |
| CN106485215B (en) | Face shielding detection method based on deep convolutional neural network | |
| WO2018003212A1 (en) | Object detection device and object detection method | |
| CN105718868B (en) | A kind of face detection system and method for multi-pose Face | |
| CN106960195A (en) | A kind of people counting method and device based on deep learning | |
| CN111160269A (en) | Face key point detection method and device | |
| CN108875537B (en) | Object detection method, device and system and storage medium | |
| CN110826379B (en) | Target detection method based on feature multiplexing and YOLOv3 | |
| CN110580445A (en) | Face key point detection method based on GIoU and weighted NMS improvement | |
| CN107688786A (en) | A kind of method for detecting human face based on concatenated convolutional neutral net | |
| CN110738160A (en) | human face quality evaluation method combining with human face detection | |
| CN109886128A (en) | A kind of method for detecting human face under low resolution | |
| CN109871821A (en) | The pedestrian of adaptive network recognition methods, device, equipment and storage medium again | |
| KR102476022B1 (en) | Face detection method and apparatus thereof | |
| CN110349167A (en) | A kind of image instance dividing method and device | |
| CN110879982A (en) | Crowd counting system and method | |
| Son et al. | Deep learning for rice quality classification | |
| CN109063626A (en) | Dynamic human face recognition methods and device | |
| CN110427912A (en) | A kind of method for detecting human face and its relevant apparatus based on deep learning | |
| CN109919246A (en) | Pedestrian's recognition methods again based on self-adaptive features cluster and multiple risks fusion | |
| CN108446602A (en) | A kind of device and method for Face datection |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |