CN110443132A - A kind of Face datection and the more attribute convergence analysis methods of face based on deep learning - Google Patents
A kind of Face datection and the more attribute convergence analysis methods of face based on deep learning Download PDFInfo
- Publication number
- CN110443132A CN110443132A CN201910590960.0A CN201910590960A CN110443132A CN 110443132 A CN110443132 A CN 110443132A CN 201910590960 A CN201910590960 A CN 201910590960A CN 110443132 A CN110443132 A CN 110443132A
- Authority
- CN
- China
- Prior art keywords
- face
- picture
- size
- region
- facial
- 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.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/168—Feature extraction; Face representation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Health & Medical Sciences (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Human Computer Interaction (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of Face datection based on deep learning and the multiattribute convergence analysis methods of face.This method are as follows: input picture first generates the different scale set of the picture using image pyramid, is handled using image of the full convolutional neural networks to input, and the first hierarchical network generates preliminary face candidate region；Then face detection module filters out face candidate frame from thick to thin, will repeat area using non-maxima suppression algorithm and simplifies beyond the candidate window of given threshold, determines human face region；Then human face region is amplified, is sent to facial attributive analysis network and is analyzed, obtains age, gender prediction's result；Finally Face datection result and attributive analysis result are labeled on picture and are saved, obtains visual prediction result.Present invention reduces the complexities of Face datection and the more attribute convergence analysis of face, have the advantages that structure is simple, parameter is few, practical.
Description
Technical field
The present invention relates to machine learning, depth learning technology field, especially a kind of Face datection based on deep learning
And the more attribute convergence analysis methods of face.
Background technique
Currently, deep learning has become the very active project of research, especially in terms of human face analysis.Such as
Modern big data era, as the computing capability of hardware greatly enhances, deep learning can more play the advantage of its own, analyze
Significant increase is obtained in precision.The experiment that many carries out in large-scale data is it has been shown that by obtained by deep learning
Character representation can show good performance in fields such as natural language processing, computer visions.
Face character refer in facial image imply some attribute informations, as the age of people, gender, hair style, adornment,
Head pose etc..Face character analysis, which refers to, to be analyzed and is identified to face character according to facial image.It is limited to current
Shooting condition, that there are resolution ratio is low for some pictures under road traffic scene, facial size is small, face obscures, blocks face etc.
Problem, how to carry out face character analysis to such picture is still the problem in computer vision field.
Summary of the invention
That the purpose of the present invention is to provide a kind of image resolution requirements is low, analysis accuracy rate is high, parameter is few, practical
Face datection based on deep learning and the more attribute convergence analysis methods of face.
The technical solution for realizing the aim of the invention is as follows: a kind of Face datection and the more attributes of face based on deep learning
Convergence analysis method, comprising the following steps:
Step 1, input picture generate the different scale set of the picture using image pyramid, using full convolutional Neural
Network handles the image of input, and the first hierarchical network generates preliminary face candidate region；
Step 2, face detection module filter out face candidate frame according to preliminary face candidate region from thick to thin, benefit
It will repeat area with non-maxima suppression algorithm to simplify beyond the candidate window of given threshold, deletion overlapping area, which is greater than, to be set
The candidate window for determining threshold value, determines human face region；
Step 3 amplifies human face region, is sent to facial attributive analysis network and is analyzed, obtains age, gender prediction
As a result；
Face datection result and attributive analysis result are labeled on picture and are saved by step 4, are obtained visual
Prediction result.
Further, input picture described in step 1 generates the different scale set of the picture using image pyramid,
It is handled using image of the full convolutional neural networks to input, the first hierarchical network generates preliminary face candidate region, tool
Body is as follows:
Step 1.1, input picture, having a size of (h, w)；
Step 1.2, the different size set that the picture is generated using image pyramid, according to the minimum facial size of setting
Zoom ratio factor between facial size net_face_size that min_face_size, the network layer can detect, pyramidal layer,
Input picture, which is zoomed to the network layer, to detect suitable size:
(h_n, w_n)=(h*net_face_size/min_face_size* (factor^n), h*net_face_size/
Min_face_size* (factor^n),
Wherein, n is the pyramid number of plies, is started counting from 0, until being less than net_face_ compared with decimal in (h_n.w_n)
Size is handled using image of the full convolutional neural networks to input, and the first hierarchical network generates preliminary face candidate area
Domain.
Further, face detection module described in step 2 filters out people according to preliminary face candidate region from thick to thin
Face candidate frame will repeat area using non-maxima suppression algorithm and simplify beyond the candidate frame of given threshold, deletes overlapping
The candidate window that area is greater than given threshold determines human face region, specific as follows:
Step 2.1 is rolled up using human face region differentiation, the recurrence of face bounding box and Face detection three cascade multitasks
Product neural network, training face detection module；
Step 2.2, using trained face detection module Stepwise Screening face candidate frame, obtain final result.
Further, human face region is amplified described in step 3, is sent to facial attributive analysis network and is analyzed, is obtained
To age, gender prediction as a result, specific as follows:
The face that step 3.1, input face detection module are predicted；
Step 3.2, the picture region that will be delivered to facial attributive analysis network expand, if a left side for protoplast's face zone boundary frame
Upper, lower right coordinate is (x1,y1), (x2,y2), then in former bounding box picture height h, width w are as follows:
Coordinate according to former bounding box improves, the practical upper left for intercepting picture region, lower right coordinate (x '1,y′1)、
(x′2,y′2) are as follows:
Increase effective information with this, includes the incoming facial attributive analysis network of information of ear, hair style by face periphery；
Step 3.3 carries out human face analysis using convolutional layer, and the convolution kernel size of three convolutional layers is 7*7,5*5,3* respectively
3, the excitation function used is PReLU, and by primary maximum pond after each convolutional layer, convolutional layer generates multiple features
Two layers of the feature in front and back is locally tied, reduces the size of characteristic pattern by face；
Step 3.4, the face characteristic extracted according to two full articulamentums, determine that the face belongs to the general of every age class
Rate and belong to male, women probability；
Step 3.5, lower probability more of all categories obtain face age, gender prediction's result.
Further, facial attributive analysis network described in step 3 selects the shooting picture under real scene as face
The data set of portion's attributive analysis.
Further, facial attributive analysis network described in step 3 is trained using online difficult example mining algorithm,
According to the difficult example ratio and this batch of total number of samples set, crucial sample number, carries out to losing in this lot sample sheet needed for calculating
Descending arrangement, extracts the crucial sample of respective numbers, forms small quantities of sample, be trained.
Compared with prior art, the present invention its remarkable advantage are as follows: (1) model prediction accuracy and network structure between the two
It is balanced, reduces image resolution requirement；(2) complexity of Face datection and the more attribute convergence analysis of face is reduced
Degree, has the advantages that structure is simple, parameter is few, practical.
Detailed description of the invention
Fig. 1 is the process signal the present invention is based on the Face datection of deep learning and the more attribute convergence analysis methods of face
Figure.
Fig. 2 is the flow diagram of face detection module in the present invention.
Fig. 3 is the flow diagram of septum reset attributive analysis of the present invention.
Fig. 4 is the prediction result schematic diagram of picture under actual scene in the embodiment of the present invention.
Specific embodiment
The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
In conjunction with Fig. 1, a kind of Face datection and the more attribute convergence analysis methods of face based on deep learning of the present invention is special
Sign is, comprising the following steps:
Step 1, input picture generate the different scale set of the picture using image pyramid, using full convolutional Neural
Network handles the image of input, and the first hierarchical network generates preliminary face candidate region, specific as follows in conjunction with Fig. 2:
Step 1.1, input picture, having a size of (h, w)；
Step 1.2, the different size set that the picture is generated using image pyramid, according to the minimum facial size of setting
Zoom ratio factor between facial size net_face_size that min_face_size, the network layer can detect, pyramidal layer,
Input picture, which is zoomed to the network layer, to detect suitable size:
(h_n, w_n)=(h*net_face_size/min_face_size* (factor^n), h*net_face_size/
Min_face_size* (factor^n),
Wherein, n is the pyramid number of plies, is started counting from 0, until being less than net_face_ compared with decimal in (h_n.w_n)
Size is handled using image of the full convolutional neural networks to input, and the first hierarchical network generates preliminary face candidate area
Domain.
Step 2, face detection module filter out face candidate frame according to preliminary face candidate region from thick to thin, benefit
It will repeat area with non-maxima suppression algorithm to simplify beyond the candidate window of given threshold, deletion overlapping area, which is greater than, to be set
The candidate window for determining threshold value, determines human face region, specific as follows in conjunction with Fig. 2:
Step 2.1 is rolled up using human face region differentiation, the recurrence of face bounding box and Face detection three cascade multitasks
Product neural network, training face detection module；
Step 2.2, using trained face detection module Stepwise Screening face candidate frame, obtain final result.
Step 3 amplifies human face region, is sent to facial attributive analysis network and is analyzed, obtains age, gender prediction
As a result, in conjunction with Fig. 3, it is specific as follows:
The face that step 3.1, input face detection module are predicted；
Step 3.2, the picture region that will be delivered to facial attributive analysis network expand, if a left side for protoplast's face zone boundary frame
Upper, lower right coordinate is (x1,y1), (x2,y2), then in former bounding box picture height h, width w are as follows:
Coordinate according to former bounding box improves, the practical upper left for intercepting picture region, lower right coordinate (x '1,y′1)、
(x′2,y′2) are as follows:
Increase effective information with this, includes the incoming facial attributive analysis network of information of ear, hair style by face periphery；
Step 3.3 carries out human face analysis using convolutional layer, and the convolution kernel size of three convolutional layers is 7*7,5*5,3* respectively
3, the excitation function used is PReLU, and by primary maximum pond after each convolutional layer, convolutional layer generates multiple features
Two layers of the feature in front and back is locally tied, reduces the size of characteristic pattern by face；
Step 3.4, the face characteristic extracted according to two full articulamentums, determine that the face belongs to the general of every age class
Rate and belong to male, women probability；
Step 3.5, lower probability more of all categories obtain face age, gender prediction's result.
Face datection result and attributive analysis result are labeled on picture and are saved by step 4, are obtained visual
Prediction result.
Further, facial attributive analysis network described in step 3 selects the shooting picture under real scene as face
The data set of portion's attributive analysis.
Further, facial attributive analysis network described in step 3 is trained using online difficult example mining algorithm,
According to the difficult example ratio and this batch of total number of samples set, crucial sample number, carries out to losing in this lot sample sheet needed for calculating
Descending arrangement, extracts the crucial sample of respective numbers, forms small quantities of sample, be trained, to increase the specific gravity of difficult example sample, by force
Change the specific aim of model.
Embodiment 1
In conjunction with Fig. 4, a specific embodiment of the invention, comprising the following steps:
Step 1, input picture generate the different scale set of the picture using image pyramid, using full convolutional Neural
Network handles the image of input, and the first hierarchical network generates preliminary face candidate region；
Step 2, face detection module filter out face candidate frame from thick to thin, will be repeated using non-maxima suppression algorithm
Area is simplified beyond the candidate window of given threshold, determines human face region；
Step 3 amplifies human face region, comprising will be enlarged by model into effective pictorial information such as hair lengths, ear etc.
Face picture after enclosing is sent to facial attributive analysis network and is analyzed, and obtains age, gender prediction's result；
Face datection result and attributive analysis result are labeled on picture and are saved by step 4, are obtained visual
Prediction result is convenient for practical application.
Specific testing result (to protect privacy, blurs face in test data) as shown in Figure 4, in test chart
In piece, rectangle frame outlines human face region, and the Gender on rectangle frame represents prediction of the present invention to the face gender, and M is male,
F is women, and the Age on rectangle frame represents prediction of the present invention to the face age class, behind institute with data represent the corresponding age
Section shares 8 kinds of possibility: 0-2,4-6,8-13,15-20,25-32,38-43,48-53,60-.Although testing the picture light used
Lines part is poor, and resolution ratio is low, and the result of model prediction is still more bonded true.As can be seen that the present invention is based on depth
The Face datection and the more attribute convergence analysis methods of face of habit, are put down between the two in model prediction accuracy and network structure
Weighing apparatus, reduces image resolution requirement, and reduces the complexity of Face datection and the more attribute convergence analysis of face, has knot
The advantage that structure is simple, parameter is few, practical.
Claims (6)
1. a kind of Face datection and the more attribute convergence analysis methods of face based on deep learning, which is characterized in that including following
Step:
Step 1, input picture generate the different scale set of the picture using image pyramid, using full convolutional neural networks
The image of input is handled, the first hierarchical network generates preliminary face candidate region；
Step 2, face detection module filter out face candidate frame, utilization is non-according to preliminary face candidate region from thick to thin
Maximum restrainable algorithms will repeat area and simplify beyond the candidate window of given threshold, delete overlapping area and are greater than setting threshold
The candidate window of value, determines human face region；
Step 3 amplifies human face region, is sent to facial attributive analysis network and is analyzed, and obtains age, gender prediction's knot
Fruit；
Face datection result and attributive analysis result are labeled on picture and are saved by step 4, are visually predicted
As a result.
2. the Face datection and the multiattribute convergence analysis method of face according to claim 1 based on deep learning,
It is characterized in that, input picture described in step 1, the different scale set of the picture is generated using image pyramid, using full volume
Product neural network handles the image of input, and the first hierarchical network generates preliminary face candidate region, specific as follows:
Step 1.1, input picture, having a size of (h, w)；
Step 1.2, the different size set that the picture is generated using image pyramid, according to the minimum facial size min_ of setting
Zoom ratio factor between facial size net_face_size that face_size, the network layer can detect, pyramidal layer, will be defeated
Suitable size can be detected to the network layer by entering image scaling:
(h_n, w_n)=(h*net_face_size/min_face_size* (factor^n), h*net_face_size/min_
Face_size* (factor^n),
Wherein, n is the pyramid number of plies, is started counting from 0, until being less than net_face_size compared with decimal in (h_n.w_n), is adopted
It is handled with image of the full convolutional neural networks to input, the first hierarchical network generates preliminary face candidate region.
3. the Face datection and the multiattribute convergence analysis method of face according to claim 1 based on deep learning,
It is characterized in that, face detection module described in step 2 filters out face candidate according to preliminary face candidate region from thick to thin
Frame will repeat area using non-maxima suppression algorithm and simplify beyond the candidate frame of given threshold, it is big to delete overlapping area
Human face region is determined in the candidate window of given threshold, specific as follows:
Step 2.1 utilizes human face region differentiation, the recurrence of face bounding box and Face detection three cascade multitask convolution minds
Through network, training face detection module；
Step 2.2, using trained face detection module Stepwise Screening face candidate frame, obtain final result.
4. the Face datection and the multiattribute convergence analysis method of face according to claim 1 based on deep learning,
It is characterized in that, amplifies human face region described in step 3, be sent to facial attributive analysis network and analyzed, obtain age, property
Other prediction result, specific as follows:
The face that step 3.1, input face detection module are predicted；
Step 3.2, the picture region that will be delivered to facial attributive analysis network expand, if the upper left of protoplast's face zone boundary frame,
Lower right coordinate is (x1,y1), (x2,y2), then in former bounding box picture height h, width w are as follows:
Coordinate according to former bounding box improves, the practical upper left for intercepting picture region, lower right coordinate (x '1,y′1)、(x′2,
y′2) are as follows:
Increase effective information with this, includes the incoming facial attributive analysis network of information of ear, hair style by face periphery；
Step 3.3 carries out human face analysis using convolutional layer, and the convolution kernel size of three convolutional layers is 7*7,5*5,3*3 respectively, is adopted
Excitation function is PReLU, and by primary maximum pond after each convolutional layer, convolutional layer generates multiple characteristic faces, will
The feature that two layers of front and back is locally tied, and reduces the size of characteristic pattern；
Step 3.4, the face characteristic extracted according to two full articulamentums, determine the face belong to every age class probability and
Belong to male, women probability；
Step 3.5, lower probability more of all categories obtain face age, gender prediction's result.
5. the Face datection and the multiattribute convergence analysis method of face according to claim 1 based on deep learning,
It is characterized in that, facial attributive analysis network described in step 3, selects the shooting picture under real scene as facial attribute point
The data set of analysis.
6. the Face datection and the multiattribute convergence analysis method of face according to claim 1 based on deep learning,
It is characterized in that, facial attributive analysis network described in step 3, is trained using online difficult example mining algorithm, according to setting
Good difficult example ratio and this batch of total number of samples, crucial sample number needed for calculating carry out descending arrangement to loss in this lot sample sheet,
The crucial sample for extracting respective numbers, forms small quantities of sample, is trained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910590960.0A CN110443132A (en) | 2019-07-02 | 2019-07-02 | A kind of Face datection and the more attribute convergence analysis methods of face based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910590960.0A CN110443132A (en) | 2019-07-02 | 2019-07-02 | A kind of Face datection and the more attribute convergence analysis methods of face based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110443132A true CN110443132A (en) | 2019-11-12 |
Family
ID=68428543
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910590960.0A Withdrawn CN110443132A (en) | 2019-07-02 | 2019-07-02 | A kind of Face datection and the more attribute convergence analysis methods of face based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110443132A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111062426A (en) * | 2019-12-11 | 2020-04-24 | 北京金山云网络技术有限公司 | Method, device, electronic equipment and medium for establishing training set |
| CN111091109A (en) * | 2019-12-24 | 2020-05-01 | 厦门瑞为信息技术有限公司 | Method, system and equipment for predicting age and gender based on face image |
| CN111626193A (en) * | 2020-05-26 | 2020-09-04 | 北京嘀嘀无限科技发展有限公司 | Face recognition method, face recognition device and readable storage medium |
| CN111914628A (en) * | 2020-06-19 | 2020-11-10 | 北京百度网讯科技有限公司 | Training method and device of face recognition model |
| CN114782400A (en) * | 2022-05-17 | 2022-07-22 | 东风本田发动机有限公司 | Method, device, equipment, medium and program product for detecting slag point of metal material |
-
2019
- 2019-07-02 CN CN201910590960.0A patent/CN110443132A/en not_active Withdrawn
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111062426A (en) * | 2019-12-11 | 2020-04-24 | 北京金山云网络技术有限公司 | Method, device, electronic equipment and medium for establishing training set |
| CN111091109A (en) * | 2019-12-24 | 2020-05-01 | 厦门瑞为信息技术有限公司 | Method, system and equipment for predicting age and gender based on face image |
| CN111091109B (en) * | 2019-12-24 | 2023-04-07 | 厦门瑞为信息技术有限公司 | Method, system and equipment for predicting age and gender based on face image |
| CN111626193A (en) * | 2020-05-26 | 2020-09-04 | 北京嘀嘀无限科技发展有限公司 | Face recognition method, face recognition device and readable storage medium |
| CN111914628A (en) * | 2020-06-19 | 2020-11-10 | 北京百度网讯科技有限公司 | Training method and device of face recognition model |
| CN111914628B (en) * | 2020-06-19 | 2023-06-20 | 北京百度网讯科技有限公司 | Training method and device of face recognition model |
| CN114782400A (en) * | 2022-05-17 | 2022-07-22 | 东风本田发动机有限公司 | Method, device, equipment, medium and program product for detecting slag point of metal material |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110443132A (en) | A kind of Face datection and the more attribute convergence analysis methods of face based on deep learning | |
| CN109101930B (en) | Crowd counting method and system | |
| CN106874894B (en) | Human body target detection method based on regional full convolution neural network | |
| CN107292256B (en) | Auxiliary task-based deep convolution wavelet neural network expression recognition method | |
| CN107358262B (en) | High-resolution image classification method and classification device | |
| WO2020119661A1 (en) | Target detection method and device and pedestrian detection method and system | |
| WO2022134655A1 (en) | End-to-end video action detection and positioning system | |
| CN109919013A (en) | Method for detecting human face and device in video image based on deep learning | |
| CN110020606A (en) | A kind of crowd density estimation method based on multiple dimensioned convolutional neural networks | |
| CN108898145A (en) | A kind of image well-marked target detection method of combination deep learning | |
| CN108416307A (en) | A kind of Aerial Images road surface crack detection method, device and equipment | |
| CN110852941B (en) | Neural network-based two-dimensional virtual fitting method | |
| US10540568B2 (en) | System and method for coarse-to-fine video object segmentation and re-composition | |
| CN111401293B (en) | Gesture recognition method based on Head lightweight Mask scanning R-CNN | |
| CN108416314B (en) | Picture important face detection method | |
| CN110348357A (en) | A kind of fast target detection method based on depth convolutional neural networks | |
| CN109635812B (en) | The example dividing method and device of image | |
| CN111696080B (en) | Face fraud detection method, system and storage medium based on static texture | |
| CN112528913A (en) | Grit particulate matter particle size detection analytic system based on image | |
| CN104484890A (en) | Video target tracking method based on compound sparse model | |
| CN109635811A (en) | The image analysis method of spatial plant | |
| CN110969171A (en) | Image classification model, method and application based on improved convolutional neural network | |
| CN111027497A (en) | Weak and small target rapid detection method based on high-resolution optical remote sensing image | |
| CN113888501B (en) | Attention positioning network-based reference-free image quality evaluation method | |
| CN107506792A (en) | A kind of semi-supervised notable method for checking object |
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 | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20191112 |
|
| WW01 | Invention patent application withdrawn after publication |