CN110458133A - Lightweight method for detecting human face based on production confrontation network - Google Patents

Lightweight method for detecting human face based on production confrontation network Download PDF

Info

Publication number
CN110458133A
CN110458133A CN201910762210.7A CN201910762210A CN110458133A CN 110458133 A CN110458133 A CN 110458133A CN 201910762210 A CN201910762210 A CN 201910762210A CN 110458133 A CN110458133 A CN 110458133A
Authority
CN
China
Prior art keywords
face
lightweight
model
network
confrontation network
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
Application number
CN201910762210.7A
Other languages
Chinese (zh)
Inventor
程建
刘济樾
周晓晔
蒋林枫
刘畅
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Electronic Science and Technology of China
Original Assignee
University of Electronic Science and Technology of China
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by University of Electronic Science and Technology of China filed Critical University of Electronic Science and Technology of China
Priority to CN201910762210.7A priority Critical patent/CN110458133A/en
Publication of CN110458133A publication Critical patent/CN110458133A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Biomedical Technology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Human Computer Interaction (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The present invention relates to the lightweight method for detecting human face based on production confrontation network, comprising: A. constructs Face datection data set；B. according to the real-time high-precision algorithm of target detection based on key point, the lightweight Face datection model based on production confrontation network is established；C. the lightweight Face datection model is trained, and according to loss, carries out stochastic gradient descent, update model parameter；D. a face picture to be detected is inputted, a propagated forward is carried out by trained lightweight Face datection model, exports the Face datection picture containing markup information.The present invention can be more easily generated clearer face in the detection process, and the speed and accuracy of detection is thus greatly improved.

Description

Lightweight method for detecting human face based on production confrontation network
Technical field
It is especially the lightweight Face datection side based on production confrontation network the present invention relates to method for detecting human face Method.
Background technique
With the appearance of deep learning, the target detection based on deep learning has obtained quantum jump.But due to depth Degree study needs powerful calculation power support, so industrially landing is difficult.Face datection is developed so far, gradually commercially Change, many Face datection algorithms have been deployed in commercial product, but Face datection task presently, there are the problem of include: Scale, posture, block, expression, appearance, illumination etc. have highly variable；Performance best face detector can be divided at present Two classes: 1) using two stages inspection policies based on the RPN (the full convolutional network of depth) applied to Faster-RCNN, and RPN generates high The candidate region (region proposal) of quality, is then further confirmed that by Fast-RCNN detector；2) based on single-lens A stage of detector SSD, the thought for being detached from RPN are directly predicted to return frame (bbox) and confidence level (confidence)。
It has been generated since Ian in 2014 is proposed based on the image that confrontation generates network, extensive concern is obtained, for face The problems such as blocking, can be taken based on the method that confrontation generates network generate it is relatively good go to block face, thus more accurately It is detected.But progress in general, is blocked in the sample to some difficult detections, such as face, the face of small resolution ratio at present When face generates, the problem that generally existing speed is slow, accuracy is low, and currently without highly effective settling mode.
Summary of the invention
The present invention provides a kind of lightweight method for detecting human face based on production confrontation network, have in the detection process Conducive to clearer face is generated, to increase the accuracy of detection.
The present invention is based on the lightweight method for detecting human face of production confrontation network, comprising:
A. Face datection data set is constructed, including in two disclosed data sets of Winder_face and CelabA Data, wherein Winder_face provides Face datection data, and is labeled with the coordinate information of real human face frame, CelabA provides the data of 5 key points；
B. it according to the real-time high-precision algorithm of target detection (CornerNet_Lite) based on key point, establishes based on generation The lightweight Face datection model of formula confrontation network；
C. the lightweight Face datection model is trained, and according to loss, carries out stochastic gradient descent, update Model parameter；
D. a face picture to be detected is inputted, a forward direction is carried out by trained lightweight Face datection model It propagates, exports the Face datection picture containing markup information.
Specifically, in step B, when establishing lightweight Face datection model, using human face target to be detected as one Point estimates the central point for returning frame to treat by the way of estimating key point.Then to face to be detected in recurrence frame Central point establishes the lightweight Face datection model according to the real-time high-precision algorithm of target detection based on key point.
It further, include: to go to block production pair by face in the lightweight Face datection model that step B is established The face for being restored to the face being blocked during Face datection that anti-network is established goes to block model (this part ginseng With the training of lightweight Face datection model), and distinguish that production confrontation network is surpassed by the good face of pre-training by small face Resolution model, the small face for carrying out deblurring recovery to the small face of the low resolution during Face datection distinguish model (this Pre-training part is partly belonged to, the training of lightweight Face datection model is not involved in).
It include ResNet (Residual Neural Network) specifically, going to block in model in the face The bottleneck layer (bottleneck) of neural network, each bottleneck layer have 21 × 1 convolution sums, 13 × 3 convolution, and using permanent Mappings are waited to reduce transmission error.
Preferably, the structure that the face goes to block in model is divided into 3 grades, and wherein the first order has 3 bottleneck layers, the second level There are 3 bottleneck layers, and using maximum pond, reduces the input dimension of input feature vector figure, the third level includes 1 bottleneck layer, and Using from attention mechanism module, improves face and go to block the ability that model captures contextual information.
Specifically, distinguishing in model in the small face, have 3 × 3 convolutional layers being of five storeys, and by by 3 different expansions The multiple dimensioned discrimination module that the empty convolutional layer of rate size is constituted, differentiates the face of different size scale.
Further, include: in step C
C1. to the lightweight Face datection model initialization, the parameter including setting lightweight Face datection model Initial value, learning rate etc.；
C2. face picture to be detected is input in the lightweight Face datection model, and exports band end to end There is the face picture of detection block；
C3. the total losses error during lightweight Face datection model training is denoted as Face datection loss function and right The sum of anti-loss function carries out error back propagation by random tonsure descent algorithm according to total losses error, updates lightweight The parameter of Face datection model.
Specifically, Face datection loss function described in step C3 are as follows:
Wherein, X is the recommendation frame of object, FCIt (X) is the class output for recommending frame X, C indicates classification, Lsoftmax() is to intersect Entropy loss (softmax), LbboxFor frame recurrence, FLIt (X) is the bounding box output of prediction, L is the posting really marked,The loss of posting for illustrating the bounding box of prediction and really marking.
Specifically, confrontation loss function described in step C3 are as follows:
Wherein, G makes a living into network, and D is to differentiate network, and x is output image, Pdata(x) true data distribution is indicated,Indicate output image x about Pdata(x) expectation, D (x) are indicated when inputting x, differentiate the output of network D, G (x) it indicates when inputting x, generates the output of network G.
Therefore, the total losses error of lightweight Face datection model in the training process is denoted as loss=LF+λLadv, λ is Measure the scale factor of loss ratio；Error is carried out using stochastic gradient descent algorithm further according to total losses error loss reversely to pass It broadcasts, updates model parameter, obtain trained lightweight Face datection model.
The present invention is based on the lightweight method for detecting human face of production confrontation network, are gone using production confrontation network generation The face blocked, reduction generates clearer face, and then increases the reliability of Face datection.By based on key point When real-time high-precision algorithm of target detection (CornerNet_Lite) establishes detection model, since CornerNet_Lite algorithm is Determine the position of detection object by the coordinate in the detection detection block upper left corner and the lower right corner, and the target in the upper left corner and the lower right corner Usually again not on face to be detected, therefore the present invention is clicked through using the thought that central point detects come the center to target face Row detection, ensure that the light weight of detection model is efficient in this way, in turn ensures the accuracy rate of detection.Especially to small resolution ratio people Face, face, which the sample for being difficult to detect such as block and be able to carry out good face, to be generated, so that the speed of detection be greatly improved With the precision of Face datection.
Specific embodiment with reference to embodiments is described in further detail above content of the invention again. But the range that this should not be interpreted as to the above-mentioned theme of the present invention is only limitted to example below.Think not departing from the above-mentioned technology of the present invention In the case of thinking, the various replacements or change made according to ordinary skill knowledge and customary means should all be included in this hair In bright range.
Detailed description of the invention
Fig. 1 is the flow chart that the lightweight method for detecting human face of network is fought the present invention is based on production.
Specific embodiment
The present invention is based on the lightweight method for detecting human face of production confrontation network as shown in Figure 1, comprising:
A. Face datection data set is constructed, including in two disclosed data sets of Winder_face and CelabA Data, wherein Winder_face provides Face datection data, and is labeled with the coordinate information of real human face frame, CelabA provides the data of 5 key points.
B. it according to the real-time high-precision algorithm of target detection (CornerNet_Lite) based on key point, establishes based on generation The lightweight Face datection model of formula confrontation network.When establishing lightweight Face datection model, by human face target to be detected Treat as a point, the central point for returning frame is estimated by the way of estimating key point.Then to be checked in frame to returning The central point for surveying face establishes the lightweight Face datection mould according to the real-time high-precision algorithm of target detection based on key point Type.
Therefore, CornetNet_Lite algorithm is the method detected based on key point, CornerNet_Lite algorithm It is the combination of two kinds of effective variants of CornerNet algorithm: CornerNet_Saccade algorithm and CornerNet_Squeeze Algorithm.Wherein CornerNet_Saccade algorithm is eliminated using attention mechanism and is thoroughly located to all pixels of image The needs of reason, and introduce the CornerNet_Squeeze of new compact backbone framework.Both variants solve effectively jointly Two crucial use-cases in target detection: it is improved efficiency in the case where not sacrificing precision, and improves the accurate of Real time Efficiency Property.
By in the zonule around the possible human face target position in CornerNet_Saccade algorithm detection image Target.It predicted using the complete image after diminution pay attention to try hard to thick bounding box, both propose possible object position It sets.Then, CornerNet_Saccade algorithm detects target by detecting the region centered on high-resolution.It may be used also To be improved efficiency by the maximum target positional number for controlling each image procossing.
CornerNet_Squeeze algorithm and pixel subset is absorbed in reduce the CornerNet_saccade for the treatment of capacity Algorithm is compared, and CornerNet_Squeeze algorithm is using MobileNetV3small version as backbone network.Specific network knot Structure is as shown in table 1:
Table 1:
Input Operation Exp size #out SE NL s
2242×3 Conv2d,3×3 - 16 - HS 2
1122×16 bneck,3×3 16 16 √ RE 2
562×16 bneck,3×3 72 24 - RE 2
282×24 bneck,3×3 88 24 - RE 1
282×24 bneck,5×5 96 40 √ HS 2
142×40 bneck,5×5 240 40 √ HS 1
142×40 bneck,5×5 240 40 √ HS 1
142×40 bneck,5×5 120 48 √ HS 1
142×48 bneck,5×5 144 48 √ HS 1
142×48 bneck,5×5 288 96 √ HS 2
72×96 bneck,5×5 576 96 √ HS 1
72×96 bneck,5×5 576 96 √ HS 1
72×96 Conv2d,1×1 - 576 √ HS 1
72×576 Pool,7×7 - - - - 1
12×576 Conv2d 1 × 1, NBN - 1280 - HS 1
12×1280 Conv2d 1 × 1, NBN - k - - 1
Wherein: Exp size indicates the size of output channel, and #out indicates that output channel number, SE are indicated whether using compression Excitation module, NL indicate that the nonlinear function used is HS module or RS module, and s indicates step-length.RE in HL is indicated ReLU nonlinear activation function, HS activation primitive (h-swish [x]) indicate are as follows:
Wherein x indicates input feature vector figure.
It include 2 models in the lightweight Face datection model of foundation, one is to go to block production by face The face for being restored to the face being blocked during Face datection that confrontation network is established goes to block model (this part Participate in the training of lightweight Face datection model), the other is distinguishing that production confrontation network is good by pre-training by small face Human face super-resolution model, the small face for carrying out deblurring recovery to the small face of the low resolution during Face datection distinguish mould Type (this partly belongs to pre-training part, is not involved in the training of lightweight Face datection model).
It wherein goes to block in model in the face, includes ResNet (Residual Neural Network) mind Bottleneck layer (bottleneck) through network, each bottleneck layer have 21 × 1 convolution sums, 13 × 3 convolution, and using identical Mapping is to reduce transmission error.And it is divided into 3 grades in the structure that face is gone to block in model, wherein the first order has 3 bottleneck layers, There are 3 bottleneck layers in the second level, and using maximum pond, reduces the input dimension of input feature vector figure, the third level includes 1 bottleneck Layer, and using from attention mechanism module, it improves face and goes to block the ability that model captures contextual information.
Distinguish in model have 3 × 3 convolutional layers being of five storeys, and by by 3 different expansion rate sizes in the small face Empty convolutional layer constitute multiple dimensioned discrimination module, the face of different size scale is differentiated.
C. optimize lightweight Face datection model:
C1. to the lightweight Face datection model initialization, the parameter including setting lightweight Face datection model Initial value, learning rate etc.；
C2. face picture to be detected is input in the lightweight Face datection model, and exports band end to end There is the face picture of detection block；
C3. the total losses error during lightweight Face datection model training is denoted as Face datection loss function and right The sum of anti-loss function carries out error back propagation by random tonsure descent algorithm according to total losses error, updates lightweight The parameter of Face datection model.
Specifically, Face datection loss function described in step C3 are as follows:
Wherein, X is the recommendation frame of object, FCIt (X) is the class output for recommending frame X, C indicates classification, Lsoftmax() is to intersect Entropy loss (softmax), bg are background, LbboxFor frame recurrence, FLIt (X) is the bounding box output of prediction, L is really marked Posting,The loss of posting for illustrating the bounding box of prediction and really marking.
Specifically, confrontation loss function described in step C3 are as follows:
Wherein, G makes a living into network, and D is to differentiate network, and x is output image, Pdata(x) true data distribution is indicated,Indicate output image x about Pdata(x) expectation, D (x) are indicated when inputting x, differentiate the output of network D, G (x) it indicates when inputting x, generates the output of network G.
Therefore, the total losses error of lightweight Face datection model in the training process is denoted as loss=LF+λLadv, λ is Measure the scale factor of loss ratio；Error is carried out using stochastic gradient descent algorithm further according to total losses error loss reversely to pass It broadcasts, updates model parameter, obtain trained lightweight Face datection model.
D. a face picture to be detected is inputted, a forward direction is carried out by trained lightweight Face datection model It propagates, exports the Face datection picture containing markup information.

Claims (9)

1. based on the lightweight method for detecting human face of production confrontation network, feature includes:
A. Face datection data set is constructed；
B. according to the real-time high-precision algorithm of target detection based on key point, the lightweight people based on production confrontation network is established Face detection model；
C. the lightweight Face datection model is trained, and according to loss, carries out stochastic gradient descent, more new model Parameter；
D. a face picture to be detected is inputted, a forward direction is carried out by trained lightweight Face datection model and is passed It broadcasts, exports the Face datection picture containing markup information.
2. the lightweight method for detecting human face as described in claim 1 based on production confrontation network, it is characterized in that: step B In, the lightweight people is established according to the real-time high-precision algorithm of target detection based on key point to the central point of face to be detected Face detection model.
3. the lightweight method for detecting human face as described in claim 1 based on production confrontation network, it is characterized in that: in step It include: to go to block that production fights that network establishes for face by face in the lightweight Face datection model that B is established The face that the face being blocked in detection process is restored goes to block model, and distinguishes that production confrontation network leads to by small face The good human face super-resolution model of pre-training is crossed, deblurring recovery is carried out to the small face of the low resolution during Face datection Small face distinguish model.
4. the lightweight method for detecting human face as claimed in claim 3 based on production confrontation network, it is characterized in that: described Face go to block in model, include the bottleneck layer of ResNet neural network, each bottleneck layer has 21 × 1 convolution sums 1 A 3 × 3 convolution, and transmission error is reduced using identical mapping.
5. the lightweight method for detecting human face as claimed in claim 4 based on production confrontation network, it is characterized in that: the people The structure that face goes to block in model is divided into 3 grades, and wherein the first order has 3 bottleneck layers, and there are 3 bottleneck layers in the second level, and uses Maximum pond reduces the input dimension of input feature vector figure, and the third level includes 1 bottleneck layer, and is used from attention mechanism module, Face is improved to go to block the ability that model captures contextual information.
6. the lightweight method for detecting human face as claimed in claim 3 based on production confrontation network, it is characterized in that: described Small face distinguish in model have 3 × 3 convolutional layers being of five storeys, and pass through by the empty convolutional layer structure of 3 different expansion rate sizes At multiple dimensioned discrimination module, the face of different size scale is differentiated.
7. the lightweight method for detecting human face based on production confrontation network as described in one of claim 1 to 6, feature Are as follows: include: in step C
C1. to the lightweight Face datection model initialization；
C2. face picture to be detected is input in the lightweight Face datection model, and output end to end is with inspection Survey the face picture of frame；
C3. the total losses error during lightweight Face datection model training is denoted as Face datection loss function and to damage-retardation The sum of function is lost, error back propagation is carried out by random tonsure descent algorithm according to total losses error, updates lightweight face The parameter of detection model.
8. the lightweight method for detecting human face as claimed in claim 7 based on production confrontation network, it is characterized in that: step C3 The Face datection loss function are as follows:
Wherein, X is the recommendation frame of object, FCIt (X) is the class output for recommending frame X, C indicates classification, Lsoftmax() is cross entropy damage It is background, L that mistake, which is bg,bboxFor frame recurrence, FLIt (X) is the bounding box output of prediction, L is the posting really marked.
9. the lightweight method for detecting human face as claimed in claim 7 based on production confrontation network, it is characterized in that: step C3 The confrontation loss function are as follows:
Wherein, G makes a living into network, and D is to differentiate network, and x is output image, Pdata(x) true data distribution is indicated,Indicate output image x about Pdata(x) expectation, D (x) are indicated when inputting x, differentiate the output of network D, G (x) it indicates when inputting x, generates the output of network G.
CN201910762210.7A 2019-08-19 2019-08-19 Lightweight method for detecting human face based on production confrontation network Pending CN110458133A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910762210.7A CN110458133A (en) 2019-08-19 2019-08-19 Lightweight method for detecting human face based on production confrontation network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910762210.7A CN110458133A (en) 2019-08-19 2019-08-19 Lightweight method for detecting human face based on production confrontation network

Publications (1)

Publication Number Publication Date
CN110458133A true CN110458133A (en) 2019-11-15

Family

ID=68487404

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910762210.7A Pending CN110458133A (en) 2019-08-19 2019-08-19 Lightweight method for detecting human face based on production confrontation network

Country Status (1)

Country Link
CN (1) CN110458133A (en)

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111428575A (en) * 2020-03-02 2020-07-17 武汉大学 Tracking method for fuzzy target based on twin network
CN111553227A (en) * 2020-04-21 2020-08-18 东南大学 Lightweight face detection method based on task guidance
CN111582141A (en) * 2020-04-30 2020-08-25 京东方科技集团股份有限公司 Face recognition model training method, face recognition method and device
CN111768342A (en) * 2020-09-03 2020-10-13 之江实验室 Human face super-resolution method based on attention mechanism and multi-stage feedback supervision
CN112686913A (en) * 2021-01-11 2021-04-20 天津大学 Object boundary detection and object segmentation model based on boundary attention consistency
CN112801069A (en) * 2021-04-14 2021-05-14 四川翼飞视科技有限公司 Face key feature point detection device, method and storage medium
CN114331904A (en) * 2021-12-31 2022-04-12 电子科技大学 Face shielding identification method
CN115311241A (en) * 2022-08-16 2022-11-08 天地（常州）自动化股份有限公司 Coal mine down-hole person detection method based on image fusion and feature enhancement

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180075581A1 (en) * 2016-09-15 2018-03-15 Twitter, Inc. Super resolution using a generative adversarial network
CN107862270A (en) * 2017-10-31 2018-03-30 深圳云天励飞技术有限公司 Face classification device training method, method for detecting human face and device, electronic equipment
CN109886167A (en) * 2019-02-01 2019-06-14 中国科学院信息工程研究所 One kind blocking face identification method and device

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180075581A1 (en) * 2016-09-15 2018-03-15 Twitter, Inc. Super resolution using a generative adversarial network
CN107862270A (en) * 2017-10-31 2018-03-30 深圳云天励飞技术有限公司 Face classification device training method, method for detecting human face and device, electronic equipment
CN109886167A (en) * 2019-02-01 2019-06-14 中国科学院信息工程研究所 One kind blocking face identification method and device

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
HEI LAW ET AL.: ""CornerNet: Detecting Objects as Paired Keypoints"", 《INTERNATIONAL JOURNAL OF COMPUTER VISION (2020)》 *
IAN J. GOODFELLOW ET AL.: ""Generative Adversarial Nets"", 《ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS》 *
贾洁: ""基于生成对抗网络的人脸超分辨率重建及识别"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 *

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111428575A (en) * 2020-03-02 2020-07-17 武汉大学 Tracking method for fuzzy target based on twin network
CN111428575B (en) * 2020-03-02 2022-03-04 武汉大学 Tracking method for fuzzy target based on twin network
CN111553227A (en) * 2020-04-21 2020-08-18 东南大学 Lightweight face detection method based on task guidance
WO2021218899A1 (en) * 2020-04-30 2021-11-04 京东方科技集团股份有限公司 Method for training facial recognition model, and method and apparatus for facial recognition
CN111582141A (en) * 2020-04-30 2020-08-25 京东方科技集团股份有限公司 Face recognition model training method, face recognition method and device
CN111768342A (en) * 2020-09-03 2020-10-13 之江实验室 Human face super-resolution method based on attention mechanism and multi-stage feedback supervision
CN112686913A (en) * 2021-01-11 2021-04-20 天津大学 Object boundary detection and object segmentation model based on boundary attention consistency
CN112686913B (en) * 2021-01-11 2022-06-10 天津大学 Object boundary detection and object segmentation model based on boundary attention consistency
CN112801069A (en) * 2021-04-14 2021-05-14 四川翼飞视科技有限公司 Face key feature point detection device, method and storage medium
CN114331904A (en) * 2021-12-31 2022-04-12 电子科技大学 Face shielding identification method
CN114331904B (en) * 2021-12-31 2023-08-08 电子科技大学 Face shielding recognition method
CN115311241A (en) * 2022-08-16 2022-11-08 天地（常州）自动化股份有限公司 Coal mine down-hole person detection method based on image fusion and feature enhancement
CN115311241B (en) * 2022-08-16 2024-04-23 天地（常州）自动化股份有限公司 Underground coal mine pedestrian detection method based on image fusion and feature enhancement

Similar Documents

Publication Publication Date Title
CN110458133A (en) Lightweight method for detecting human face based on production confrontation network
CN112801164B (en) Training method, device, equipment and storage medium of target detection model
CN108182456B (en) Target detection model based on deep learning and training method thereof
CN106796716B (en) For providing the device and method of super-resolution for low-resolution image
CN109727246A (en) Comparative learning image quality evaluation method based on twin network
CN106650699A (en) CNN-based face detection method and device
CN108765506A (en) Compression method based on successively network binaryzation
CN109558902A (en) A kind of fast target detection method
CN108229591A (en) Neural network adaptive training method and apparatus, equipment, program and storage medium
CN109784283A (en) Based on the Remote Sensing Target extracting method under scene Recognition task
CN111161254A (en) Bone age prediction method
CN109711401A (en) A kind of Method for text detection in natural scene image based on Faster Rcnn
CN112580733B (en) Classification model training method, device, equipment and storage medium
CN106991666A (en) A kind of disease geo-radar image recognition methods suitable for many size pictorial informations
CN106203350B (en) A kind of across the scale tracking of moving target and device
CN113205041B (en) Structured information extraction method, device, equipment and storage medium
CN113642431A (en) Training method and device of target detection model, electronic equipment and storage medium
CN108550163A (en) Moving target detecting method in a kind of complex background scene
WO2024032010A1 (en) Transfer learning strategy-based real-time few-shot object detection method
CN114463759A (en) Lightweight character detection method and device based on anchor-frame-free algorithm
CN112836692B (en) Method, apparatus, device and medium for processing image
CN108351962A (en) Object detection with adaptivity channel characteristics
CN108229494A (en) network training method, processing method, device, storage medium and electronic equipment
CN109977834A (en) The method and apparatus divided manpower from depth image and interact object
CN111222534B (en) Single-shot multi-frame detector optimization method based on bidirectional feature fusion and more balanced L1 loss

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

Application publication date: 20191115

RJ01 Rejection of invention patent application after publication