CN109409210A - A kind of method for detecting human face and system based on SSD frame - Google Patents
A kind of method for detecting human face and system based on SSD frame Download PDFInfo
- Publication number
- CN109409210A CN109409210A CN201811058101.9A CN201811058101A CN109409210A CN 109409210 A CN109409210 A CN 109409210A CN 201811058101 A CN201811058101 A CN 201811058101A CN 109409210 A CN109409210 A CN 109409210A
- Authority
- CN
- China
- Prior art keywords
- face
- anchor point
- facial image
- size
- sample
- 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
- G06V40/161—Detection; Localisation; Normalisation
-
- 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
Abstract
The embodiment of the invention provides a kind of method for detecting human face and system based on SSD frame, this method comprises: marking anchor point to target facial image；Based on the SSD frame convolutional neural networks after training, the target facial image for being labeled with anchor point is detected, gets target face frame, the SSD frame convolutional neural networks are obtained by being labeled with the sample facial image training of anchor point.The anchor point matching algorithm that the embodiment of the present invention passes through the adaptive face size distribution of setting, face in every facial image is matched into most suitable anchor point according to facial size size, to generate trained human face region, SSD frame convolutional neural networks are trained up, the performance of Face datection is improved.
Description
Technical field
The present embodiments relate to deep learning field more particularly to a kind of method for detecting human face based on SSD frame and
System.
Background technique
In recent years, with the fast development of the new and high technologies such as computer, human face detection tech neural network based is wide
It is general to apply the convolutional neural networks for being based particularly on SSD training frame in fields such as photography, monitoring and gate inhibitions that there is detection speed
Fastly, it detects the big object aspect high feature of accuracy rate and is paid close attention to by people.
Due to being limited to low-level feature abstract deficiency, the convolutional neural networks based on SSD frame are various sizes of in detection
Effect when face is poor, especially small size face (size that small size face refers to is in 50px*50px face below)
Aspect performance is relatively poor, in the limited situation of training data, the information of mining data cannot be trained, be caused well
Training is insufficient, reduces the performance of the convolutional neural networks monitoring face of SSD frame.
Therefore, it needs a kind of to solve the above problems based on the method for detecting human face of SSD frame and system now.
Summary of the invention
In view of the problems of the existing technology, the embodiment of the invention provides a kind of Face datection sides based on SSD frame
Method and system.
In a first aspect, the embodiment of the invention provides a kind of method for detecting human face based on SSD frame, comprising:
Anchor point is marked to target facial image；
Based on the SSD frame convolutional neural networks after training, the target facial image for being labeled with anchor point is detected, is obtained
Target face frame is got, the SSD frame convolutional neural networks are obtained by being labeled with the sample facial image training of anchor point.
Second aspect, the embodiment of the invention provides a kind of face detection systems based on SSD frame, comprising:
Anchor point labeling module, for marking anchor point to target facial image；
Face detection module, for based on the SSD frame convolutional neural networks after training, to the target person for being labeled with anchor point
Face image is detected, and gets target face frame, and the SSD frame convolutional neural networks are by being labeled with the sample graph of anchor point
As training obtains.
The third aspect the embodiment of the invention provides a kind of electronic equipment, including memory, processor and is stored in storage
On device and the computer program that can run on a processor, realize that such as first aspect mentions when the processor executes described program
The step of method of confession.
Fourth aspect, the embodiment of the invention provides a kind of non-transient computer readable storage mediums, are stored thereon with meter
Calculation machine program is realized as provided by first aspect when the computer program is executed by processor the step of method.
A kind of method for detecting human face and system based on SSD frame provided in an embodiment of the present invention, it is adaptive by being arranged
The anchor point matching algorithm of face size distribution, the face in every facial image is most suitable according to the matching of facial size size
Anchor point trains up SSD frame convolutional neural networks, improves Face datection to generate trained human face region
Performance.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is this hair
Bright some embodiments for those of ordinary skill in the art without creative efforts, can be with root
Other attached drawings are obtained according to these attached drawings.
Fig. 1 is the flow diagram of the method for detecting human face provided in an embodiment of the present invention based on SSD frame；
Fig. 2 is the training flow diagram of SSD frame convolutional neural networks provided in an embodiment of the present invention；
Fig. 3 is the structural schematic diagram of the face detection system provided in an embodiment of the present invention based on SSD frame；
Fig. 4 is electronic devices structure schematic diagram provided in an embodiment of the present invention.
Specific embodiment
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with the embodiment of the present invention
In attached drawing, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described embodiment is
A part of the embodiment of the present invention, instead of all the embodiments.Based on the embodiments of the present invention, those of ordinary skill in the art
Every other embodiment obtained without creative efforts, shall fall within the protection scope of the present invention.
Face datection, which refers in the picture given by certain computerized algorithm to any one width, detects people therein
Face region, and the position of human face region is marked out to the relevant technologies come in picture.SSD frame convolutional neural networks are in face
Have detection speed fast in detection, detects the big object aspect high feature of accuracy rate and paid close attention to by people, but be limited to low layer
Feature extraction deficiency causes its performance in terms of detecting small size face relatively poor, common in the art based on SSD's
Face datection frame main problem are as follows: low-level feature network structure is relatively easy, cannot sufficiently excavate small size face characteristic, make
At in terms of small size face detection performance it is poor (size that small size face refers to is in 50px*50px face below)；Instruction
Anchor point cannot be matched well by practicing data, caused to train insufficient, training data cannot be sufficiently excavated, especially in training data
In limited situation, the information of mining data it cannot be trained well；Due to face training sample small-medium size face, in
The specific gravity of equidimension face and large scale face is not necessarily balanced, and various sizes of face is caused to be to the contribution of loss function
It is different, it is unfavorable for optimizing network parameter.
Therefore, how to effectively improve performance of the SSD frame convolutional neural networks in terms of detecting small size face and mention
The problem of recall rate and accuracy rate of high entire detector are present urgent need to resolve.
It includes face or personage's body that sample facial image and target facial image described in the embodiment of the present invention, which refer to,
The image of body also can be used as the embodiment of the present invention even if face or character physical in image is imperfect or only includes part
In facial image.In addition, sample facial image and target facial image only to it in SSD frame convolutional neural networks into
Made difference when row training or application, actual nature is the same between the two, is facial image, the embodiment of the present invention is not
It is specifically limited.
Fig. 1 is the flow diagram of the method for detecting human face provided in an embodiment of the present invention based on SSD frame, such as Fig. 1 institute
Show, the embodiment of the invention provides a kind of method for detecting human face based on SSD frame, comprising:
Step 101, anchor point is marked to target facial image；
Step 102, based on the SSD frame convolutional neural networks after training, to be labeled with the target facial image of anchor point into
Row detection, gets target face frame, and the SSD frame convolutional neural networks are instructed by being labeled with the sample facial image of anchor point
It gets.
In embodiments of the present invention, first by step 101, anchor point is marked to target facial image, in SSD frame convolution
In neural network, for face characteristic in significantly more efficient extraction training data, anchor point is marked to each sample facial image.
Step 202, based on the SSD frame convolutional neural networks after training, the target facial image for being labeled with anchor point is detected, is obtained
Target face frame is got, the SSD frame convolutional neural networks are obtained by being labeled with the sample facial image training of anchor point,
In training process, for feature extraction layer, SSD frame convolutional neural networks can be raw in each of sample facial image position
At the anchor point of certain amount, for matching the face in training data, the face being only matched to by anchor point could treat as positive sample
This (demarcates the face of anchor point), and the region without the matched anchor point delineation of face can be identified as negative sample (background), so
Positive sample and negative sample region can extract feature respectively afterwards, are sent into loss function and classified and returned.
Therefore, the matched validity of anchor point directly affects the effect of classification, influences final detection performance.For example, if
One anchor point is matched to a human face region many comprising background, will carry more redundancy noise when classification,
Influence classifying quality.In the ideal case, if each anchor point is just matched to human face region, and ambient noise is not included
When, classifying quality can be optimal.The embodiment of the present invention chooses corresponding ruler according to the facial size being distributed in sample facial image
Transformation ratio is spent, scaling processing is carried out to sample facial image, then the sample facial image after scaling is taken just at random
Square region, and the anchor point for being square the matching corresponding size of the face in region is labeled, and obtains the sample for being labeled with anchor point
This facial image, scales and a series of processing such as go mean value finally by Random Level overturning, picture, obtains for SSD frame
The training sample set of frame convolutional neural networks.
The embodiment of the present invention, will be in every facial image by the anchor point matching algorithm of the adaptive face size distribution of setting
Face most suitable anchor point matched according to facial size size make SSD frame convolutional Neural to generate trained human face region
Network is trained up, and the performance of Face datection is improved.
On the basis of the above embodiments, Fig. 2 is the training of SSD frame convolutional neural networks provided in an embodiment of the present invention
Flow diagram, as shown in Fig. 2, the SSD frame convolutional neural networks are obtained by following steps training:
Step 201, random change of scale is carried out to sample facial image, obtains first sample facial image；
Step 202, square area is taken at random to the first sample facial image, if the square area contains
Face, the then anchor point for matching corresponding size are labeled, and obtain the sample facial image for being labeled with anchor point；
Step 203, Random Level overturning is carried out to the sample facial image for being labeled with anchor point, obtains the second sample people
Face image；
Step 204, by the size scaling of the second sample facial image to network inputs size and mean value is subtracted, obtained
Third sample facial image is trained the SSD frame convolutional neural networks by the third sample facial image.
In embodiments of the present invention, due to the face of training data different scale greatly in the picture of shooting, data
Distribution may unevenly, such as shoot parade picture small-medium size face accounting it is relatively large, and shone in front, party takes pictures
Etc. large scales face accounting it is larger.Therefore the face for different scale is needed to do anchor point matching respectively.In order to increase data
Diversity enables the small size face in original image to be matched to relatively large anchor point, and large scale face can also be matched to phase
To lesser anchor point, original image is carried out random scaling by this patent, then takes off region from the picture after scaling
With anchor point, the small size face amplified in this way can also be matched to big anchor point, and large scale face can also be matched to small anchor point, thus
So that various sizes of anchor point has been obtained training, improves the ability to express of network.
Random change of scale is carried out to sample facial image first, specific as follows:
S1, the face number for being less than 32px and the size greater than 256px in a sample facial image is calculated separately, and
The face sum of all sizes, is denoted as num_small, num_large, num_total respectively in this sample facial image；
If S2, num_small/num_total are greater than certain threshold value (such as 0.8), illustrate this sample face figure
Small size face as in is occupied than great, then makees the random place of size constancy or enhanced processing to sample facial image
Reason；
If the condition of S3, S2 are unsatisfactory for, the value of num_large/num_total is calculated, if it is greater than certain threshold value (example
As 0.8), illustrates occupying than great for large scale face in this figure, then size constancy or diminution are made to sample facial image
The random process of processing；
If the condition in S4, S2 and S3 is not satisfied, illustrate the face distribution phase of various sizes in sample facial image
To random process that is uniform, then making size constancy to sample facial image, zoom in or out；
First sample facial image is obtained by above-mentioned random scaling step, then in first sample facial image
Square area is taken at random, the square area without face is given up, is contained until in the square area taken at random
Face, then corresponding anchor point is matched for the square area containing face, as the sample facial image for being labeled with anchor point.Then
The sample facial image that anchor point will be labeled with carries out Random Level overturning, obtains the second sample facial image, then by the second sample
Facial image zooms to network inputs size, for example, zooming to input size 640px*640px, wherein zooming parameter includes
INTER_LINEAR, INTER_CUBIC, INTER_AREA, INTER_NEAREST, INTER_LANCZOS4, each scaling ginseng
Number randomly selects, and the sample facial image after scaling is finally subtracted mean value, obtains the instruction for SSD frame convolutional neural networks
Practice sample set.
The embodiment of the present invention, will be in every facial image by the anchor point matching algorithm of the adaptive face size distribution of setting
Face most suitable anchor point matched according to facial size size make SSD frame convolutional Neural to generate trained human face region
Network is trained up, and the performance of Face datection is improved.
On the basis of the above embodiments, random change of scale is carried out to sample facial image described, obtains the first sample
Before this facial image, comprising:
The sample facial image is pre-processed, the pretreatment includes contrast, brightness plus white noise or tone
Processing.
In embodiments of the present invention, various picture processing, including contrast, brightness plus white noise are carried out to original image
Or the processing of tone, while each pretreated method being randomly selected by probability 0.5.In entire preprocessing process
In, main purpose is to eliminate the redundancy in sample facial image, filters out interference and noise, restores necessary true letter
Breath, to increase the detectability of relevant information, and then the extraction of Enhanced feature, the segmentation of image, matching and identification can
By property, the quality of data is improved for subsequent processing as far as possible.
It is on the basis of the above embodiments, described that random change of scale is carried out to sample facial image, comprising:
Change of scale ratio is set according to the facial size distribution situation in the sample facial image, for the sample
Facial image carries out random change of scale, and the change of scale ratio is 0.5,1 or 2.
It in embodiments of the present invention, is 0.5,1 or 2 by change of scale ratio setting, corresponding operation is by sample
The length and width of this facial image reduce 0.5 times, remain unchanged, twice of length and width amplification, according to various sizes of in sample facial image
Face is distributed accounting situation, chooses corresponding change of scale ratio.For example, the small size face in sample facial image occupies
When specific gravity is larger, change of scale ratio is randomly selected from 1.0,2.0, if large scale face occupy specific gravity it is larger when, scale become
Ratio is changed to randomly select from 0.5,1.0, if the face distribution of the various sizes in sample facial image is relatively uniform, scale
Transformation ratio is randomly selected from 0.5,1.0,2.0.
The embodiment of the present invention ensure that each sample facial image can be with by the way that corresponding change of scale ratio is arranged
Most suitable anchor point is matched, and the training effectiveness of SSD frame convolutional neural networks is improved.
On the basis of the above embodiments, described that square area is taken at random to the first sample facial image, if
The square area contains face, then the anchor point for matching corresponding size is labeled, and obtains the sample face for being labeled with anchor point
Image, comprising:
If the square area contains face, and the size of face is more than or equal to the ratio between the size of anchor point to be matched
Preset threshold, the then anchor point for matching corresponding size are labeled；
If the square area contains face, and the ratio between the size of face and the size of anchor point to be matched are less than described
Preset threshold then gives up the square area；
If the square area does not contain face, give up the square area.
In embodiments of the present invention, one piece of square area is taken at random to first sample facial image first, if this is just
Square region is free of face, then gives up this square area, then proceed to take first sample facial image at random, until scratching
Contain face in the square human face region taken, (default is the width and network inputs size for then calculating the square area
640) ratio, then be multiplied respectively with the longest edge of face in the square area, it obtains corresponding under network inputs size
Then facial size S chooses and carries out anchor point mark with this facial size S and immediate anchor point size, for example, anchor point size
Be 16,32,64,128,256 and 512, choose while, it is necessary to assure facial size S be less than anchor point size, and with most
The ratio between close anchor point size is more than or equal to preset threshold, and in embodiments of the present invention, setting preset threshold is 0.9, if face
When size S and immediate anchor point size carry out anchor point mark, the ratio between facial size S and immediate anchor point size are less than default
Threshold value then gives up the square area.In addition, if the operation for take square area to first sample facial image reaches
After certain number, for example, if the square area taken each time does not comply with requirement, just will after taking 50 times
First sample facial image continues subsequent processing.It should be noted that the embodiment of the present invention includes one with square area
A face is illustrated, and when in square area including multiple faces, is successively carried out above-mentioned step to multiple faces and is carried out
Processing, if all people's face does not meet preset threshold in the square area, gives up the square area, specific steps
It repeats no more.
The embodiment of the present invention is by matching adaptive face size to the square area taken in sample facial image
Anchor point guarantees that the face in square area can be matched with most suitable anchor point, to sufficiently excavate training data, makes
SSD frame convolutional neural networks are trained up, and the performance of Face datection is improved.
On the basis of the various embodiments described above, square region is taken at random to the first sample facial image described
Domain, if the square area contains face, the anchor point for matching corresponding size is labeled, and obtains the sample for being labeled with anchor point
After facial image, comprising:
By the facial image for including in the sample facial image for being labeled with anchor point by small size facial image, medium ruler
Very little facial image and large scale facial image are classified；
Face penalty values after calculating separately facial image size classes, and according to formula:
Loss=L1+L2+L3+L4
Calculating final face penalty values, wherein L1 is small size face penalty values, and L2 is medium size face penalty values,
L3 is large scale face penalty values, and L4 is all size face's penalty values；
The sample facial image for being labeled with anchor point is optimized according to the final face penalty values, for described
SSD frame convolutional neural networks are trained.
In embodiments of the present invention, each feature extraction layer is each responsible for the face of one size of detection, such as
Conv3_3 is responsible for detecting the face of 16px.Since small size face is unfavorable for detecting high-rise size is relatively small,
Layer detects small size face (16px, 32px) SSD frame convolutional neural networks according to this from low to high, medium face (64px,
128px), large scale face (256px, 512px), but due in low layer characteristic pattern it is relatively large sized, small size face
It is again relatively small, therefore when calculating loss function, the positive sample Limited Number that anchor point is matched to, and negative sample institute accounting
High, the positive and negative proportional imbalance of weight.In order to alleviate this problem, OHEM algorithm is used in SSD frame, only chooses 3 times of positive sample numbers
Purpose negative sample calculates loss function, then calculates the penalty values of the positive negative sample of the extraction of all feature extraction layers.But by
The size of face is that have a certain range in normal picture, and common is that small size face and medium size face are in the majority, in this way
Result in the shared penalty values ratio of large scale face relatively small, and small size face and the face of medium size size
Penalty values proportion is higher, and will lead to large scale face in this way cannot train well, and effect is relatively poor.In order to solve
The unbalanced problem of the face of different scale proportion in loss function, the embodiment of the present invention is by small size face, medium
The penalty values of size face and large scale face are respectively calculated, and are then summed up again, while all size faces are added
Penalty values, network is optimized using this tetrameric penalty values as final penalty values.Specifically, in small size face class
In not, using the anchor point data for matching small size face as positive sample, it is not matched to the region of face anchor point as negative sample,
Medium size face and large scale face are also made equally to handle.
The embodiment of the present invention calculates separately face penalty values to various sizes of face, then sums up again, adds simultaneously
The face penalty values for entering the size carry out training sample using this tetrameric face penalty values as final penalty values
Optimization, it is unbalanced to avoid various sizes of face penalty values proportion in loss function, leads to be unable to balance optimizing not
The problem of with size face, improves the recall rate and accuracy of Face datection.
On the basis of the various embodiments described above, further, the face calculated separately after facial image size classes
Penalty values specifically include:
Various sizes of face penalty values, formula in the sample image of anchor point are labeled with described in calculating are as follows:
Wherein, Lk({pk,i},{tk,i) lost for the face of the sample image for being labeled with anchor point under kth kind size
Value, i is the label of an anchor point, pk,iIt is in the face penalty values of the sample image for being labeled with anchor point under kth kind size
I-th of anchor point is predicted as the probability of face,It is the label of mark, wherein the label containing face is 1, is free of face label
It is 0, tk,iIt is four position prediction coordinate vectors,It is four face anchor point coordinate vectors,It is Softmax loss
Function,It is Smooth L1 loss function, NclsFor the anchor point number of face and Background matching, NregFor face
The number for the anchor point matched, λ are the weight for matching two kinds of loss functions, default value 1；
After face penalty values in getting the sample facial image under different sizes, the final face is calculated
Penalty values, formula are as follows:
Wherein, λkIt is the corresponding weighted value of the sorted four kinds of face penalty values of face scale, default value 1.It needs
Bright, the classification function and regression function of the embodiment of the present invention are not limited to Softmax loss function and Smooth L1 damage
Function is lost, other classification functions and regression function also can operate with the embodiment of the present invention.Therefore, the embodiment of the present invention is not to loss
Function makees specific limit.
The embodiment of the present invention calculates separately penalty values to the face of different scale, then resulting penalty values are weighted place
Reason, avoids the face penalty values of different scale in loss function, due to proportion it is unbalanced caused by cannot be balanced
The problem of optimizing different scale face, improves the detection performance of SSD convolutional neural networks.
Fig. 3 is the structural schematic diagram of the face detection system provided in an embodiment of the present invention based on SSD frame, such as Fig. 3 institute
Show, the embodiment of the invention provides a kind of face detection systems based on SSD frame, including anchor point labeling module 301 and face
Detection module 302, wherein anchor point labeling module 301 is used to mark anchor point to target facial image；Face detection module 302 is used
In based on the SSD frame convolutional neural networks after training, the target facial image for being labeled with anchor point is detected, mesh is got
Face frame is marked, the SSD frame convolutional neural networks are obtained by being labeled with the sample image training of anchor point.
In embodiments of the present invention, for face characteristic, anchor point labeling module 301 in significantly more efficient extraction training data
Anchor point is marked to each sample facial image in SSD frame convolutional neural networks is input to.When SSD frame convolutional neural networks
After the completion of training, anchor point is marked to target facial image by anchor point labeling module 301, after the completion of anchor point mark, face inspection
It surveys module 302 to detect the target facial image for being labeled with anchor point, to get target face frame.SSD frame convolution
Neural network is in training process, and for feature extraction layer, SSD frame convolutional neural networks can be every in sample facial image
One position generates the anchor point of certain amount, and for matching the face in training data, the face being only matched to by anchor point is
Negative sample can be identified as positive sample (demarcating the face of anchor point), the region without the matched anchor point delineation of face
(background), then positive sample and negative sample region can extract feature respectively, are sent into loss function and classified and returned.
Therefore, the matched validity of anchor point directly affects the effect of classification, influences final detection performance.In ideal situation
Under, if each anchor point is just matched to human face region, and when not including ambient noise, classifying quality can be optimal.This
Inventive embodiments choose corresponding scale transformation ratio according to the facial size being distributed in sample facial image, to sample facial image
Scaling processing is carried out, then the sample facial image after scaling is taken into square area at random, and be square in region
Face matching corresponding size anchor point be labeled, the sample facial image for being labeled with anchor point is obtained, finally by random water
Flat overturning, picture scaling and a series of processing such as go mean value, obtain the training sample for SSD frame convolutional neural networks
Collection.
The embodiment of the present invention, will be in every facial image by the anchor point matching algorithm of the adaptive face size distribution of setting
Face most suitable anchor point matched according to facial size size improve Face datection to generate trained human face region
Accuracy rate and recall rate.
Fig. 4 is electronic devices structure schematic diagram provided in an embodiment of the present invention, as shown in figure 4, the electronic equipment can wrap
It includes: processor (processor) 401, communication interface (Communications Interface) 402, memory (memory)
403 and communication bus 404, wherein processor 401, communication interface 402, memory 403 are completed mutually by communication bus 404
Between communication.Processor 401 can call the logical order in memory 403, to execute following method: to target facial image
Mark anchor point；Based on the SSD frame convolutional neural networks after training, the target facial image for being labeled with anchor point is detected,
Target face frame is got, the SSD frame convolutional neural networks are obtained by being labeled with the sample facial image training of anchor point.
In addition, the logical order in above-mentioned memory 403 can be realized by way of SFU software functional unit and conduct
Independent product when selling or using, can store in a computer readable storage medium.Based on this understanding, originally
Substantially the part of the part that contributes to existing technology or the technical solution can be in other words for the technical solution of invention
The form of software product embodies, which is stored in a storage medium, including some instructions to
So that a computer equipment (can be personal computer, server or the network equipment etc.) executes each implementation of the present invention
The all or part of the steps of example the method.And storage medium above-mentioned include: USB flash disk, mobile hard disk, read-only memory (ROM,
Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic or disk etc. it is various
It can store the medium of program code.
The embodiment of the present invention discloses a kind of computer program product, and the computer program product is non-transient including being stored in
Computer program on computer readable storage medium, the computer program include program instruction, when described program instructs quilt
When computer executes, computer is able to carry out method provided by above-mentioned each method embodiment, for example, to target face figure
As mark anchor point；Based on the SSD frame convolutional neural networks after training, the target facial image for being labeled with anchor point is examined
It surveys, gets target face frame, the SSD frame convolutional neural networks are trained by the sample facial image for being labeled with anchor point
It arrives.
The embodiment of the present invention provides a kind of non-transient computer readable storage medium, the non-transient computer readable storage medium
The instruction of matter storage server, the computer instruction make computer execute the face based on SSD frame provided by above-described embodiment
Detection method, for example, anchor point is marked to target facial image；Based on the SSD frame convolutional neural networks after training, to mark
The target facial image for being marked with anchor point is detected, and target face frame is got, and the SSD frame convolutional neural networks pass through mark
The sample facial image training for being marked with anchor point obtains.
Finally, it should be noted that the above embodiments are merely illustrative of the technical solutions of the present invention, rather than its limitations；Although
Present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that: it still may be used
To modify the technical solutions described in the foregoing embodiments or equivalent replacement of some of the technical features；
And these are modified or replaceed, technical solution of various embodiments of the present invention that it does not separate the essence of the corresponding technical solution spirit and
Range.
Claims (10)
1. a kind of method for detecting human face based on SSD frame characterized by comprising
Anchor point is marked to target facial image；
Based on the SSD frame convolutional neural networks after training, the target facial image for being labeled with anchor point is detected, is got
Target face frame, the SSD frame convolutional neural networks are obtained by being labeled with the sample facial image training of anchor point.
2. the method according to claim 1, wherein the SSD frame convolutional neural networks pass through following steps
Training obtains:
Random change of scale is carried out to sample facial image, obtains first sample facial image；
Square area is taken to the first sample facial image at random, if the square area contains face, is matched
The anchor point of corresponding size is labeled, and obtains the sample facial image for being labeled with anchor point；
Random Level overturning is carried out to the sample facial image for being labeled with anchor point, obtains the second sample facial image；
By the size scaling of the second sample facial image to network inputs size and mean value is subtracted, obtains third sample face
Image is trained the SSD frame convolutional neural networks by the third sample facial image.
3. according to the method described in claim 2, it is characterized in that, carrying out random scale change to sample facial image described
It changes, before obtaining first sample facial image, comprising:
The sample facial image is pre-processed, the pretreatment includes the place of contrast, brightness plus white noise or tone
Reason.
4. according to the method described in claim 2, it is characterized in that, described carry out random change of scale to sample facial image,
Include:
Change of scale ratio is set according to the facial size distribution situation in the sample facial image, for the sample face
Image carries out random change of scale, and the change of scale ratio is 0.5,1 or 2.
5. according to the method described in claim 2, it is characterized in that, described take just the first sample facial image at random
Square region, if the square area contains face, the anchor point for matching corresponding size is labeled, and obtains being labeled with anchor point
Sample facial image, comprising:
If the square area contains face, and the ratio between the size of face and the size of anchor point to be matched are more than or equal to default
Threshold value, the then anchor point for matching corresponding size are labeled；
If the square area contains face, and the ratio between the size of face and the size of anchor point to be matched are less than described preset
Threshold value then gives up the square area；
If the square area does not contain face, give up the square area.
6. according to the method described in claim 2, it is characterized in that, being taken at random described to the first sample facial image
Square area, if the square area contains face, the anchor point for matching corresponding size is labeled, and obtains being labeled with anchor
After the sample facial image of point, comprising:
By the facial image for including in the sample facial image for being labeled with anchor point by small size facial image, medium size people
Face image and large scale facial image are classified；
Face penalty values after calculating separately facial image size classes, and according to formula:
Loss=L1+L2+L3+L4
Calculate final face penalty values, wherein L1 is small size face penalty values, and L2 is medium size face penalty values, and L3 is
Large scale face penalty values, L4 are all size face's penalty values；
The sample facial image for being labeled with anchor point is optimized according to the final face penalty values, for the SSD
Frame convolutional neural networks are trained.
7. according to the method described in claim 6, it is characterized in that, the face calculated separately after facial image size classes
Penalty values specifically include:
Various sizes of face penalty values, formula in the sample image of anchor point are labeled with described in calculating are as follows:
Wherein, Lk({pk,i},{tk,i) be the sample image for being labeled with anchor point under kth kind size face penalty values, i
It is the label of an anchor point, pk,iIt is i-th in the face penalty values of the sample image for being labeled with anchor point under kth kind size
A anchor point is predicted as the probability of face,It is the label of mark, wherein the label containing face is 1, it is 0 without face label,
tk,iIt is four position prediction coordinate vectors,It is four face anchor point coordinate vectors,It is Softmax loss letter
Number,It is Smooth L1 loss function, NclsFor the anchor point number of face and Background matching, NregFor face matching
Anchor point number, λ be match two kinds of loss functions weight, default value 1；
After face penalty values in getting the sample facial image under different sizes, the final face loss is calculated
Value, formula are as follows:
Wherein, λkIt is the corresponding weighted value of the sorted four kinds of face penalty values of face scale, default value 1.
8. a kind of face detection system based on SSD frame characterized by comprising
Anchor point labeling module, for marking anchor point to target facial image；
Face detection module, for based on the SSD frame convolutional neural networks after training, to the target face figure for being labeled with anchor point
As being detected, target face frame is got, the SSD frame convolutional neural networks are instructed by being labeled with the sample image of anchor point
It gets.
9. a kind of electronic equipment including memory, processor and stores the calculating that can be run on a memory and on a processor
Machine program, which is characterized in that the processor is realized when executing described program such as any one of claim 1 to 7 the method
Step.
10. a kind of non-transient computer readable storage medium, is stored thereon with computer program, which is characterized in that the computer
It is realized when program is executed by processor such as the step of any one of claim 1 to 7 the method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811058101.9A CN109409210B (en) | 2018-09-11 | 2018-09-11 | Face detection method and system based on SSD (solid State disk) framework |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811058101.9A CN109409210B (en) | 2018-09-11 | 2018-09-11 | Face detection method and system based on SSD (solid State disk) framework |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109409210A true CN109409210A (en) | 2019-03-01 |
| CN109409210B CN109409210B (en) | 2020-11-24 |
Family
ID=65464778
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811058101.9A Active CN109409210B (en) | 2018-09-11 | 2018-09-11 | Face detection method and system based on SSD (solid State disk) framework |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109409210B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110197113A (en) * | 2019-03-28 | 2019-09-03 | 杰创智能科技股份有限公司 | A kind of method for detecting human face of high-precision anchor point matching strategy |
| CN110210314A (en) * | 2019-05-06 | 2019-09-06 | 深圳市华付信息技术有限公司 | Method for detecting human face, device, computer equipment and storage medium |
| CN110490115A (en) * | 2019-08-13 | 2019-11-22 | 北京达佳互联信息技术有限公司 | Training method, device, electronic equipment and the storage medium of Face datection model |
| CN111079707A (en) * | 2019-12-31 | 2020-04-28 | 深圳云天励飞技术有限公司 | Face detection method and related device |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107220618A (en) * | 2017-05-25 | 2017-09-29 | 中国科学院自动化研究所 | Method for detecting human face and device, computer-readable recording medium, equipment |
| CN107403141A (en) * | 2017-07-05 | 2017-11-28 | 中国科学院自动化研究所 | Method for detecting human face and device, computer-readable recording medium, equipment |
| CN108230354A (en) * | 2017-05-18 | 2018-06-29 | 深圳市商汤科技有限公司 | Target following, network training method, device, electronic equipment and storage medium |
| CN108229442A (en) * | 2018-02-07 | 2018-06-29 | 西南科技大学 | Face fast and stable detection method in image sequence based on MS-KCF |
| CN108256400A (en) * | 2016-12-29 | 2018-07-06 | 上海玄彩美科网络科技有限公司 | The method for detecting human face of SSD based on deep learning |
-
2018
- 2018-09-11 CN CN201811058101.9A patent/CN109409210B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108256400A (en) * | 2016-12-29 | 2018-07-06 | 上海玄彩美科网络科技有限公司 | The method for detecting human face of SSD based on deep learning |
| CN108230354A (en) * | 2017-05-18 | 2018-06-29 | 深圳市商汤科技有限公司 | Target following, network training method, device, electronic equipment and storage medium |
| CN107220618A (en) * | 2017-05-25 | 2017-09-29 | 中国科学院自动化研究所 | Method for detecting human face and device, computer-readable recording medium, equipment |
| CN107403141A (en) * | 2017-07-05 | 2017-11-28 | 中国科学院自动化研究所 | Method for detecting human face and device, computer-readable recording medium, equipment |
| CN108229442A (en) * | 2018-02-07 | 2018-06-29 | 西南科技大学 | Face fast and stable detection method in image sequence based on MS-KCF |
Non-Patent Citations (1)
| Title |
|---|
| SHUO YANG等: "Face Detection through Scale-Friendly Deep Convolutional Networks", 《COMPUTER VISION AND PATTERN RECOGNITION》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110197113A (en) * | 2019-03-28 | 2019-09-03 | 杰创智能科技股份有限公司 | A kind of method for detecting human face of high-precision anchor point matching strategy |
| CN110197113B (en) * | 2019-03-28 | 2021-06-04 | 杰创智能科技股份有限公司 | Face detection method of high-precision anchor point matching strategy |
| CN110210314A (en) * | 2019-05-06 | 2019-09-06 | 深圳市华付信息技术有限公司 | Method for detecting human face, device, computer equipment and storage medium |
| CN110490115A (en) * | 2019-08-13 | 2019-11-22 | 北京达佳互联信息技术有限公司 | Training method, device, electronic equipment and the storage medium of Face datection model |
| CN110490115B (en) * | 2019-08-13 | 2021-08-13 | 北京达佳互联信息技术有限公司 | Training method and device of face detection model, electronic equipment and storage medium |
| CN111079707A (en) * | 2019-12-31 | 2020-04-28 | 深圳云天励飞技术有限公司 | Face detection method and related device |
| CN111079707B (en) * | 2019-12-31 | 2023-06-13 | 深圳云天励飞技术有限公司 | Face detection method and related device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109409210B (en) | 2020-11-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Li et al. | Instance-level salient object segmentation | |
| CN109409210A (en) | A kind of method for detecting human face and system based on SSD frame | |
| Talebi et al. | Learned perceptual image enhancement | |
| CN109492638A (en) | Method for text detection, device and electronic equipment | |
| CN110298298A (en) | Target detection and the training method of target detection network, device and equipment | |
| CN109583449A (en) | Character identifying method and Related product | |
| CN107977665A (en) | The recognition methods of key message and computing device in a kind of invoice | |
| CN110097051A (en) | Image classification method, device and computer readable storage medium | |
| CN111738243B (en) | Method, device and equipment for selecting face image and storage medium | |
| CN108961279A (en) | Image processing method, device and mobile terminal | |
| CN109101897A (en) | Object detection method, system and the relevant device of underwater robot | |
| CN110263819A (en) | A kind of object detection method and device for shellfish image | |
| CN110363084A (en) | A kind of class state detection method, device, storage medium and electronics | |
| CN108846826A (en) | Object detecting method, device, image processing equipment and storage medium | |
| CN110008956A (en) | Invoice key message localization method, device, computer equipment and storage medium | |
| CN109472193A (en) | Method for detecting human face and device | |
| CN109741293A (en) | Conspicuousness detection method and device | |
| CN110490073A (en) | Object detection method, device, equipment and storage medium | |
| EP4047509A1 (en) | Facial parsing method and related devices | |
| CN110689658A (en) | Taxi bill identification method and system based on deep learning | |
| CN110288602A (en) | Come down extracting method, landslide extraction system and terminal | |
| CN106372597B (en) | CNN Vehicle Detection method based on adaptive contextual information | |
| CN110069959A (en) | A kind of method for detecting human face, device and user equipment | |
| CN110084175A (en) | A kind of object detection method, object detecting device and electronic equipment | |
| CN109426773A (en) | A kind of roads recognition method and device |
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 | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20200914 Address after: 215000 unit 2-b702, creative industry park, No. 328, Xinghu street, Suzhou Industrial Park, Suzhou City, Jiangsu Province Applicant after: SUZHOU FEISOU TECHNOLOGY Co.,Ltd. Address before: 100876 Beijing, Haidian District, 10 West Road, Beijing, 12 Beijing, North Post Science and technology exchange center, room 1216 Applicant before: BEIJING FEISOU TECHNOLOGY Co.,Ltd. |
|
| TA01 | Transfer of patent application right | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |