CN108090417A - A kind of method for detecting human face based on convolutional neural networks - Google Patents
A kind of method for detecting human face based on convolutional neural networks Download PDFInfo
- Publication number
- CN108090417A CN108090417A CN201711204234.8A CN201711204234A CN108090417A CN 108090417 A CN108090417 A CN 108090417A CN 201711204234 A CN201711204234 A CN 201711204234A CN 108090417 A CN108090417 A CN 108090417A
- Authority
- CN
- China
- Prior art keywords
- face
- face datection
- convolutional neural
- neural networks
- detecting human
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/168—Feature extraction; Face representation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06V40/172—Classification, e.g. identification
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Oral & Maxillofacial Surgery (AREA)
- General Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Human Computer Interaction (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of method for detecting human face based on convolutional neural networks, comprise the following steps:1) Face datection model is established, which uses RFCN network structures, and the RFCN network structures include the feature extraction layer of feature based fusion;2) sample set is obtained;3) the Face datection model established in step 1) is trained;4) mapping piece is treated with the Face datection model after training and carries out Face datection.Compared with prior art, the present invention have many advantages, such as accuracy rate and recall ratio it is higher, for have under complex scene it is good adapt to effect.
Description
Technical field
The present invention relates to technical field of face recognition, more particularly, to a kind of Face datection side based on convolutional neural networks
Method.
Background technology
Face datection is one and is related to the multi-field research topic such as computer vision, pattern-recognition and artificial intelligence, because
It is widely applied value in the fields such as business, medical treatment and military affairs, is always the hot spot of people's research.However, in real field
Under scape, often there is serious shielding in the face in complicated image, this brings huge challenge to Face datection, so
It is proposed it is a kind of can adapt to be still in the method for detecting human face seriously blocked research difficult point.
Document " Object Detection via Region-based Fully Convolutional Networks "
(Dai,J., Li,Y.,He,K.,Sun,J.:R-FCN:.In:30th Conference on Neural Information
Processing Systems, pp.379-387.Barcelona) a kind of target based on the full convolutional neural networks in region is disclosed
Detection method, this method basic network use ResNet101, are RFCN network structures, and sub-network is divided into Region Proposal
Network (RPN) and sorter network, overall network structure are as shown in Figure 1.The process of ResNet extractions feature maps totally 4
A stage is denoted as res1, res2, res3, res4 respectively.Pass through convolution algorithm and RPN sub-networks and classification subnet after res4
Connection.RPN sub-networks share the feature maps that ResNet is extracted with classification subnet so that the extraction of feature only need to be into
Row once-through operation, drastically increases operation efficiency.
RPN networks are used to extract region proposals, that is, possible human face region.Rpn_bbox_pred layers
Return each region compared with anchor offset.Anchor is based on the different scale for being originally inputted picture and being generated
The rectangle frame of scale and length-width ratio ratio.Each anchor values are directed to each anchor plus what rpn_bbox_pred was obtained
Offset be exactly RPN layers and need the positions of region exported.It is foreground object that rpn_cls_prob, which exports each region,
With the probability of background.Proposal layers are integrated the result of rpn_bbox_pred layers and rpn_cls_prob layers, root
It is ranked up according to prospect probability, then using non-maxima suppression non maximum supression (NMS) if algorithm obtains
Dry regions.(2000 being extracted during training, 300 are extracted during test).Fiveth stage res5 of the sorter network based on ResNet
Continue to obtain the score maps that depth is C*k*k after extracting feature.K is hyper parameter, value 3;C represents the classification of final classification
Number (includes background classes), value 2 (face | background).RFCN utilizes ROIPooling layers of Position-sensitive, to RPN
Each region that network obtains, is location-based average pooling on score maps.It is to region's
Feature is extracted in each position respectively, by drawing final result to the ballot of all positions.By RPN sub-networks and divide
The probability that class sub-network finally can obtain the position region where face and each region is face.
The object detection method in the training process, selects common data sets WIDER FACE first to be obtained as sample set
The RFCN models of pre-training on ImageNet, then start to train on ready sample set again.Finally with training after
Model carries out Face datection.
Although above-mentioned existing method can obtain certain precision, also have the following disadvantages:1st, sensitivity is blocked for face, had
It is more block in the case of detection difficult, mAP only has 0.77 on WIDER FACE;2nd, examined for smaller face or side face
It surveys bad.
The content of the invention
It is an object of the present invention to overcome the above-mentioned drawbacks of the prior art and provide one kind is based on convolutional Neural
The method for detecting human face of network.
The purpose of the present invention can be achieved through the following technical solutions:
A kind of method for detecting human face based on convolutional neural networks, comprises the following steps:
1) Face datection model is established, which uses RFCN network structures, and the RFCN network structures are included based on spy
Levy the feature extraction layer of fusion;
2) sample set is obtained;
3) the Face datection model established in step 1) is trained;
4) mapping piece is treated with the Face datection model after training and carries out Face datection.
Further, in the feature extraction layer, by the output layer of res3 with the output layer of res4 is superimposed merges.
Further, in the step 2), the sample size of sample set is more than 30,000.
Further, the training of the step 3) uses caffe frames, including:
301) pre-training is carried out to the Face datection model on ImageNet;
302) pretrained rear face detection model is trained again using the sample set.
Further, in the step 4), multiple scale detecting is carried out to the picture to be measured.
Further, the multiple scale detecting is specially:
401) the flexible processing of multiple sizes is carried out to the picture to be measured;
402) Face datection is carried out respectively to the picture obtained under each size using the Face datection model after training, obtained
Much a face testing results;
403) screening is merged to the multiple Face datection result, obtains final detection result.
Further, in the step 403), sieve is merged to the multiple Face datection result using NMS algorithms
Choosing.
Compared with prior art, the invention has the advantages that:
1) present invention establishes a follow-on Face datection model, in characteristic extraction procedure, the output to res3
Output with res4 carries out Fusion Features, and the feature that fusion obtains can be simultaneously applied two sons of RPN and sorter network
In net, the accuracy rate and recall ratio of Face datection are substantially increased.
2) present invention, using multiple scale detecting mode, can get and more be hidden when being detected to tested picture
The information of the face of gear and small resolution ratio face further improves the accuracy rate and recall ratio of Face datection.
3) sample size of sample of the present invention collection is more than 30,000, ensure that the accuracy of detection model.
4) present invention is for there is good adaptation effect under complex scene, especially for face serious shielding and the field of small face
Scape, by largely testing, accuracy rate and recall ratio reach more than 90%.
Description of the drawings
Fig. 1 is the overall network structure diagram of existing method;
Fig. 2 is existing RPN sub-network structures schematic diagram;
Fig. 3 is the schematic network structure of feature of present invention fusion;
Fig. 4 is the overall network structure diagram of the present invention;
Fig. 5 is the testing process schematic diagram of the present invention;
Fig. 6 is the detection result schematic diagram of the present invention.
Specific embodiment
The present invention is described in detail with specific embodiment below in conjunction with the accompanying drawings.The present embodiment is with technical solution of the present invention
Premised on implemented, give detailed embodiment and specific operating process, but protection scope of the present invention is not limited to
Following embodiments.
The present invention provides a kind of method for detecting human face based on convolutional neural networks, comprises the following steps:1) face is established
Detection model, the model use RFCN network structures, and the RFCN network structures include the feature extraction layer of feature based fusion;
2) sample set is obtained;3) the Face datection model established in step 1) is trained;4) with the Face datection model after training
It treats mapping piece and carries out Face datection.Accuracy rate and the higher face of recall ratio can be carried out to complex scene by the above method
Detection.
The key point of above-mentioned detection method is:
A, model structure is improved
The improvement of model structure essentially consists in network interlayer Fusion Features.In the structure of ResNet101, preceding four ranks
Section has done 4 pooling operations altogether until res4, and deeper convolution pooling networks cause each feature map's
Receptive field is bigger, and the semantic feature learnt is also more advanced.But face or smaller face for serious shielding, tool
Standby feature is inherently limited, extracts the local feature that high-level semantic feature causes its limited and is easier to lose.Namely
It says, for the object that the feature that exposes is limited, big receptive field plays the role of detection to be not so good as small receptive field.Therefore it is
The face of the small face of correct detection and serious shielding, the present invention is by the output layer of res3 and the output layer of res4 is superimposed melts
It closes, makes the feature that network learns at res4 layers that there is high-level semantics feature and rudimentary local feature simultaneously.Selection is in res4
The reason for being merged is that the feature that fusion obtains can be simultaneously applied in two subnets of RPN and sorter network.Feature
For the network structure of fusion as shown in figure 3, res4b22_relu is the output of res4, res4b22_dcov is the output knot to res4
The up-sampling of fruit makes the feature map of the feature map and res3 of res4 keep the size of same size, res3_
Scale expands the channel numbers of res3, keeps the depth with res4feature map same sizes.Due to
Deconvolution operations can only double into even number or odd number, and the operation before pooling is even number or odd number is
Indefinite, it is therefore desirable to it is operated using crop and the feature map after deconvolution is cropped to the ruler identical with res3
It is very little.Improved overall network structure is as shown in Figure 4.
B, the training stage
The first step:Makes sample
Sample set source is common data sets WIDER FACE, in total comprising 32,203 pictures, 393,703 faces
Sample.It is made according to the form of PASCAL VOC data sets, PASCAL VOC provide a whole set of mark for image recognition and calssification
The outstanding data set of standardization, therefore face sample is made by this standard.Sample stores specification:It is deposited in JPEGImages
The samples pictures for including face are put, face mesh in the details for corresponding to samples pictures and picture is stored in Annotations
Target encirclement frame coordinate, wherein face frame position mark form are made of top left co-ordinate and lower left corner coordinate, Annotation
It is stored using xml document form.
Second step:Training pattern
Caffe frames are used for the training of model.The improved RFCN moulds of the pre-training on ImageNet are obtained first
Then type starts to train on ready sample set again.Shown in trained hyper parameter table 1.
Table 1:Trained hyper parameter is set
iterations | 500000 | batch size | 1 |
base learning rate | 0.001 | k | 3 |
momentum | 0.9 | scale | 1,2,4 |
weight_decay | 0.0005 | ratio | 0.5,1,2 |
Training obtains model file face_model.caffemodel, utilizes the i.e. detectable face of the model file.
C, face is detected
For the Face datection under complex scene, it is intended that get the face being more blocked and small resolution ratio people
The information of face.Therefore when detecting, multiple dimensioned processing is carried out to picture, the picture of each scale is subjected to one-time detection,
Then screening is merged using NMS algorithms to the result of detection, so as to obtain final testing result, flow such as Fig. 5 institutes
Show.
MAP on WIRDER FACE reaches 0.897, and detection result is as shown in Figure 6.
The preferred embodiment of the present invention described in detail above.It should be appreciated that those of ordinary skill in the art without
Creative work is needed according to the present invention can to conceive and makes many modifications and variations.Therefore, all technologies in the art
Personnel are available by logical analysis, reasoning, or a limited experiment on the basis of existing technology under this invention's idea
Technical solution, all should be in the protection domain being defined in the patent claims.
Claims (7)
1. a kind of method for detecting human face based on convolutional neural networks, which is characterized in that comprise the following steps:
1) Face datection model is established, which uses RFCN network structures, and the RFCN network structures are melted including feature based
The feature extraction layer of conjunction;
2) sample set is obtained;
3) the Face datection model established in step 1) is trained;
4) mapping piece is treated with the Face datection model after training and carries out Face datection.
2. the method for detecting human face according to claim 1 based on convolutional neural networks, which is characterized in that the feature carries
Take in layer, by the output layer of res3 with the output layer of res4 is superimposed merges.
3. the detection method of raising one's hand according to claim 1 based on deep learning, which is characterized in that in the step 2),
The sample size of sample set is more than 30,000.
4. the method for detecting human face according to claim 1 based on convolutional neural networks, which is characterized in that the step 3)
Training using caffe frames, including:
301) pre-training is carried out to the Face datection model on ImageNet;
302) pretrained rear face detection model is trained again using the sample set.
5. the method for detecting human face according to claim 1 based on convolutional neural networks, which is characterized in that the step 4)
In, multiple scale detecting is carried out to the picture to be measured.
6. the method for detecting human face according to claim 5 based on convolutional neural networks, which is characterized in that described multiple dimensioned
Detection is specially:
401) the flexible processing of multiple sizes is carried out to the picture to be measured;
402) carry out Face datection respectively to the picture obtained under each size using the Face datection model after training, obtain more
A face testing result;
403) screening is merged to the multiple Face datection result, obtains final detection result.
7. the method for detecting human face according to claim 6 based on convolutional neural networks, which is characterized in that the step
403) in, screening is merged to the multiple Face datection result using NMS algorithms.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711204234.8A CN108090417A (en) | 2017-11-27 | 2017-11-27 | A kind of method for detecting human face based on convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711204234.8A CN108090417A (en) | 2017-11-27 | 2017-11-27 | A kind of method for detecting human face based on convolutional neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108090417A true CN108090417A (en) | 2018-05-29 |
Family
ID=62173195
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711204234.8A Pending CN108090417A (en) | 2017-11-27 | 2017-11-27 | A kind of method for detecting human face based on convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108090417A (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108830205A (en) * | 2018-06-04 | 2018-11-16 | 江南大学 | Based on the multiple dimensioned perception pedestrian detection method for improving full convolutional network |
CN109101899A (en) * | 2018-07-23 | 2018-12-28 | 北京飞搜科技有限公司 | A kind of method for detecting human face and system based on convolutional neural networks |
CN110070124A (en) * | 2019-04-15 | 2019-07-30 | 广州小鹏汽车科技有限公司 | A kind of image amplification method and system based on production confrontation network |
CN110189255A (en) * | 2019-05-29 | 2019-08-30 | 电子科技大学 | Method for detecting human face based on hierarchical detection |
CN110602411A (en) * | 2019-08-07 | 2019-12-20 | 深圳市华付信息技术有限公司 | Method for improving quality of face image in backlight environment |
CN110659659A (en) * | 2018-07-19 | 2020-01-07 | 清华珠三角研究院 | Method and system for intelligently identifying and early warning pests |
CN110728310A (en) * | 2019-09-27 | 2020-01-24 | 聚时科技(上海)有限公司 | Target detection model fusion method and system based on hyper-parameter optimization |
CN112686126A (en) * | 2020-12-25 | 2021-04-20 | 杭州海康威视数字技术股份有限公司 | Face modeling method and device |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106650615A (en) * | 2016-11-07 | 2017-05-10 | 深圳云天励飞技术有限公司 | Image processing method and terminal |
CN106845421A (en) * | 2017-01-22 | 2017-06-13 | 北京飞搜科技有限公司 | Face characteristic recognition methods and system based on multi-region feature and metric learning |
CN106874894A (en) * | 2017-03-28 | 2017-06-20 | 电子科技大学 | A kind of human body target detection method based on the full convolutional neural networks in region |
CN107145908A (en) * | 2017-05-08 | 2017-09-08 | 江南大学 | A kind of small target detecting method based on R FCN |
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 |
-
2017
- 2017-11-27 CN CN201711204234.8A patent/CN108090417A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106650615A (en) * | 2016-11-07 | 2017-05-10 | 深圳云天励飞技术有限公司 | Image processing method and terminal |
CN106845421A (en) * | 2017-01-22 | 2017-06-13 | 北京飞搜科技有限公司 | Face characteristic recognition methods and system based on multi-region feature and metric learning |
CN106874894A (en) * | 2017-03-28 | 2017-06-20 | 电子科技大学 | A kind of human body target detection method based on the full convolutional neural networks in region |
CN107145908A (en) * | 2017-05-08 | 2017-09-08 | 江南大学 | A kind of small target detecting method based on R FCN |
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 |
Non-Patent Citations (2)
Title |
---|
JIFENG DAI ET AL.: "R-FCN: Object Detection via Region-based Fully Convolutional Networks", 《ARXIV:1605.06409V2》 * |
LIN LIU ET AL.: "Highly Occluded Face Detection: An Improved R-FCN Approach", 《ICONIP 2017》 * |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108830205B (en) * | 2018-06-04 | 2019-06-14 | 江南大学 | Based on the multiple dimensioned perception pedestrian detection method for improving full convolutional network |
CN108830205A (en) * | 2018-06-04 | 2018-11-16 | 江南大学 | Based on the multiple dimensioned perception pedestrian detection method for improving full convolutional network |
US10977521B2 (en) | 2018-06-04 | 2021-04-13 | Jiangnan University | Multi-scale aware pedestrian detection method based on improved full convolutional network |
WO2019232836A1 (en) * | 2018-06-04 | 2019-12-12 | 江南大学 | Multi-scale sensing pedestrian detection method based on improved full convolutional network |
CN110659659A (en) * | 2018-07-19 | 2020-01-07 | 清华珠三角研究院 | Method and system for intelligently identifying and early warning pests |
CN109101899A (en) * | 2018-07-23 | 2018-12-28 | 北京飞搜科技有限公司 | A kind of method for detecting human face and system based on convolutional neural networks |
CN109101899B (en) * | 2018-07-23 | 2020-11-24 | 苏州飞搜科技有限公司 | Face detection method and system based on convolutional neural network |
CN110070124A (en) * | 2019-04-15 | 2019-07-30 | 广州小鹏汽车科技有限公司 | A kind of image amplification method and system based on production confrontation network |
CN110189255A (en) * | 2019-05-29 | 2019-08-30 | 电子科技大学 | Method for detecting human face based on hierarchical detection |
CN110602411A (en) * | 2019-08-07 | 2019-12-20 | 深圳市华付信息技术有限公司 | Method for improving quality of face image in backlight environment |
CN110728310A (en) * | 2019-09-27 | 2020-01-24 | 聚时科技(上海)有限公司 | Target detection model fusion method and system based on hyper-parameter optimization |
CN110728310B (en) * | 2019-09-27 | 2023-09-01 | 聚时科技(上海)有限公司 | Target detection model fusion method and fusion system based on super-parameter optimization |
CN112686126A (en) * | 2020-12-25 | 2021-04-20 | 杭州海康威视数字技术股份有限公司 | Face modeling method and device |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108090417A (en) | A kind of method for detecting human face based on convolutional neural networks | |
CN108961272B (en) | Method for generating skin disease image based on deep convolution countermeasure generation network | |
CN108875732B (en) | Model training and instance segmentation method, device and system and storage medium | |
CN107123123A (en) | Image segmentation quality evaluating method based on convolutional neural networks | |
CN109886179A (en) | The image partition method and system of cervical cell smear based on Mask-RCNN | |
CN104143079B (en) | The method and system of face character identification | |
CN107730507A (en) | A kind of lesion region automatic division method based on deep learning | |
CN108475350A (en) | Use the model accelerated semiconductor correlation computations based on study | |
CN109801256A (en) | A kind of image aesthetic quality appraisal procedure based on area-of-interest and global characteristics | |
CN108427912A (en) | Remote sensing image object detection method based on the study of dense target signature | |
CN107423701A (en) | The non-supervisory feature learning method and device of face based on production confrontation network | |
CN108010021A (en) | A kind of magic magiscan and method | |
CN108198179A (en) | A kind of CT medical image pulmonary nodule detection methods for generating confrontation network improvement | |
CN107871101A (en) | A kind of method for detecting human face and device | |
CN105160361A (en) | Image identification method and apparatus | |
CN107423398A (en) | Exchange method, device, storage medium and computer equipment | |
CN107066934A (en) | Tumor stomach cell image recognition decision maker, method and tumor stomach section identification decision equipment | |
CN111754453A (en) | Pulmonary tuberculosis detection method and system based on chest radiography image and storage medium | |
CN109636817A (en) | A kind of Lung neoplasm dividing method based on two-dimensional convolution neural network | |
CN104680545B (en) | There is the detection method of well-marked target in optical imagery | |
CN105608454A (en) | Text structure part detection neural network based text detection method and system | |
CN108765392A (en) | A kind of digestive endoscope lesion detection and recognition methods based on sliding window | |
CN110516517A (en) | A kind of target identification method based on multiple image, device and equipment | |
CN109801225A (en) | Face reticulate pattern stain minimizing technology based on the full convolutional neural networks of multitask | |
Mohmmad et al. | A survey machine learning based object detections in an image |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180529 |