CN107563279A - The model training method adjusted for the adaptive weighting of human body attributive classification - Google Patents
The model training method adjusted for the adaptive weighting of human body attributive classification Download PDFInfo
- Publication number
- CN107563279A CN107563279A CN201710603212.2A CN201710603212A CN107563279A CN 107563279 A CN107563279 A CN 107563279A CN 201710603212 A CN201710603212 A CN 201710603212A CN 107563279 A CN107563279 A CN 107563279A
- Authority
- CN
- China
- Prior art keywords
- task
- weight
- error
- model
- attribute
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Abstract
The invention belongs to computer visual image processing technology field, the multi task model training method specially adjusted for the adaptive weighting of human body attributive classification.The present invention proposes a kind of novel multi task model, by introducing one based on validation error size and variation tendency so as to update the algorithm of corresponding task weight, adaptively dynamically adjusts the respective weights value of each task in the training process.Specific steps include:(1)The collection of face and pedestrian's picture and the other mark of Attribute class;(2)Build deep neural network;(3)Train deep neural network;(4)Using depth network model, human body attribute forecast is carried out;The inventive method has the advantages that speed is fast, accuracy is high, robustness is good, is highly suitable for the practical applications such as the related detection of human body, identification, classification.
Description
Technical field
The invention belongs to computer visual image processing technology field, and in particular to for the adaptive of human body attributive classification
The multi task model training method of weight adjustment.
Background technology
Human body attribute analysis technology is in actual life, such as environmental safety monitoring system, traffic control monitor system etc.
Have a wide range of applications.Human body attribute is also for advanced Computer Vision Tasks such as the identification of people's body weight, clothing matchings simultaneously
Key feature, thus human body attribute analysis technology receives researchers and more and more paid close attention to.Under conditions of unrestricted, by
In the change of human posture, block, light etc. influences, and human body attributive analysis remains a larger challenge.
With the recovery of convolutional neural networks, depth multitask network on different classes of attribute by carrying out shared spy
The method that sign represents carries out the prediction of human body attribute, often in a network in addition to top for single attribute task, remaining
The parameter all properties of each layer are all shared.Such method can produce two problems:1. when two attribute task gaps are larger
When, the performance that can damage learner is changed in insufficient violence.2. multi-task learning method as can pre-set and instruct
The respective weights of each task are fixed in white silk, thus do not consider the otherness and relevance between different task.Though based on more
The learning method of business is widely used on attribute forecast, but based on two problems more than this, the learning method of multitask at present
Certain defect be present, because that can be interfered during study between multiple tasks, and how during training,
Effectively the weight of adjustment control different task also fails to propose an excellent solution.
In the model training of multitask, literary [1,2] is directed to facial critical point detection task, employs while predictive
Not, the secondary attribute such as expression, appearance task lifts the detection performance of goal task.Literary [3] propose a permission different task
Between share the convolutional neural networks of some visual information multitasks and learn face character.Literary [4] excavate Face datection and and
Interdependency between alignment lifts the performance of depth cascade multitask network.However, be directed in multi task model
The adjustment control of the weight of different task remains a problem.
To solve above urgent problem to be solved, unlike the method for existing attributive analysis, invention introduces
The algorithm of one weight that can dynamically adjust each task in the training process based on validation error, so as to propose one kind
The multi task model training method of adaptive weighting adjustment trains multitask human body attributive analysis model.
The content of the invention
It is an object of the invention to propose a kind of multi task model training method of adaptive weighting adjustment, to train more
Business human body attributive analysis model.
Weight adjustment control for different task in multi task model is a problem urgently to be resolved hurrily.In order to solve
This problem, the present invention propose a kind of novel multi task model, and introducing one in multi task model is based on validation error
The weight that can dynamically adjust each task in the training process algorithm, make adaptively to dynamically adjust in the training process
The respective weights value of each task.Specifically, the importance of a task is weighed using generalization ability, for generalization ability
Poor model sets higher weights.For each attributive analysis task, its error and error on checking collection is analyzed
Variation tendency.For the task that error is larger, show that this task has difficulty;For the task of Error Trend change greatly, table
Show that the learnability for being directed to this task model is preferable, give higher weight parameter.Pass through such weight adjustable strategies, energy
Enough so that each attribute task is sufficiently learnt, it is unlikely to over-fitting again, has highlighted the innovation of the present invention.
The multi task model training method of adaptive weighting adjustment proposed by the present invention, is comprised the following steps that:
(1)The collection of face and pedestrian's picture and the other mark of Attribute class;This method needs to collect certain original image number
According to the training for human body attributive analysis model;The attribute classification that should have corresponding human body for every pictures marks;Face category
Property include whether to wear glasses, if wear masks, if wear sunglasses, if make up, if young, hair length, whether hair
Curling, eyebrows bushed together, eyes size, if be double-edged eyelid, nose whether Gao Ting, if having double chin etc.;The attribute bag of pedestrian
Include sleeve length, lower body garment length, garment language, knapsack, handbag, upper body clothing color and lower body garment color etc.;
(2)Build deep neural network;Infrastructure network framework uses ResNet-50 frameworks;Fig. 1 is designed depth god
Structure through network;Including:Input layer, basic network, multitask weight key-course, basic network include convolutional layer, full connection
Layer, pond layer etc.;Wherein:
Input layer is responsible for receiving input;
Input picture passes through the convolutional layer conv1 of first layer, then through pond layer, and pass through 16 alternate convolution modules(Quan Lian
Connect layer)Carry out feature extraction;Pond layer is connected with after first and the convolutional layer of last, pond layer can be to adjacent region
Thresholding, which is done, to be polymerize so that the certain deformation of network tolerable, strengthens translation and the rotational invariance of learning characteristic;By convolution and
After pondization operation, the feature that extraction is obtained inputs full articulamentum, and full articulamentum is that a linear transformation is done to the feature of input,
Can be by the Projection Character of input to a more preferable subspace, so as to complete attribute forecast task;Network is finally more
Business weight key-course, it is responsible for calculating the difference value between the attribute and markup information of prediction, and is completed by backpropagation to weight
Adaptive adjustment;
(3)Train deep neural network;Multitask people is trained using the multi task model training method of adaptive weighting adjustment
Body attributive analysis model, one weight that can dynamically adjust each task in the training process based on validation error of introducing
Algorithm;Weight can change at different moments in training of each task, the change of weight is being tested according to each task
Error size and variation tendency on card collection determine, by constantly iterating to calculate and backpropagation, optimize depth network model
In parameter;
(4)Using the depth network model in step 3, human body attribute forecast task is carried out;Train and complete in above-mentioned depth model
Afterwards, for a given face or pedestrian's picture, can export in the image for the pre- of face character or pedestrian's attribute
Survey result.
The innovation of the present invention is:
1. a kind of multi task model training method of adaptive weighting adjustment is proposed to train multitask human body attributive analysis model.
Its weight parameter is independently updated according to the generalization ability of the corresponding learner of each task, solves multitask mould breakthroughly
For the weight adjustment control problem of different task in type;
2. the deep neural network for carrying out human body attributive analysis task, adds multitask weight key-course, a base is introduced
In the algorithm of the weight that can dynamically adjust each task in the training process of validation error.In end-to-end training process
In, using backpropagation mode, control is adjusted to the corresponding weight parameter of each task, to strengthen the extensive energy of model
Power.The training method updated using weight, the analysis prediction task accuracy rate of face character and human body attribute, which has, substantially to be carried
Rise.
Brief description of the drawings
Fig. 1 is the network model framework schematic diagram of human body attributive analysis.
Embodiment
The collection of step 1. face and pedestrian's picture and the other mark of Attribute class.The classification of face character has outside face
Subordinate's property and face inherent attribute composition, the external attribute of face include whether to wear glasses, if wear masks, if wear ink
Mirror, if make up etc..The inherent attribute of face can be subdivided into integrity attribute and local attribute again.Integrity attribute includes sex,
It is whether young, face value etc..In terms of local attribute then focuses on the face details of face, including whether hair length, hair crimp,
Eyebrows bushed together, eyes size, if be double-edged eyelid, nose whether Gao Ting, if having double chin etc..The attribute of pedestrian includes sleeve
Length, lower body garment length, garment language, knapsack, handbag, upper body clothing color and lower body garment color etc..
Step 2. builds deep neural network.Infrastructure network framework uses ResNet-50 frameworks, and Fig. 1 is set
The structure of the deep neural network of meter.In addition to last multitask weight key-course, full articulamentum and all pond layers, network
In every layer below be respectively connected with non-linear layer, using Relu as activation primitive, function representation is f (x)=max (0, x).
Input picture passes through the convolutional layer conv1 of first layer first, then through pond layer, and pass through 16 alternate convolution moulds
Block carries out feature extraction.Pond layer is connected with after first and the convolutional layer of last, pond layer can be to adjacent area
Value, which is done, to be polymerize so that the certain deformation of network tolerable, strengthens translation and the rotational invariance of learning characteristic.By convolution and pond
After changing operation, the feature that extraction is obtained inputs full articulamentum, and full articulamentum is that a linear transformation, energy are done to the feature of input
Enough Projection Characters by input are to a more preferable subspace, so as to complete attribute forecast task.The last of network is multitask
Weight key-course, it is responsible for calculating the difference value between the attribute and markup information of prediction.
In general, input layer is responsible for receiving input.By alternate convolutional layer, the combination of non-linear layer and pond layer,
Carry out the feature extraction of picture.Full articulamentum can be mapped the feature of acquisition.Last multitask weight key-course is born
The prediction error of calculating network is blamed, and the adaptive adjustment to weight is completed by backpropagation.
Step 3. trains deep neural network.The label information for being ready to complete face, human body picture and corresponding attribute with
Afterwards, the training of depth network is carried out.The multi task model training side adjusted with reference to accompanying drawing 1 to adaptive weighting proposed by the present invention
Method elaborates.For the training image of input after some convolutional layers and pond layer, the characteristic pattern input that extraction is obtained is more
Calculated in task weight key-course, weight can change at different moments in training of each task, the change of weight
Change is to verify the error size on collecting and variation tendency decision according to each task.The letter of network is described in detail below
Breath.
Fig. 1 shows the main frame of method.There are 3 main parts, convolutional layer, full articulamentum and multitask weight
Key-course.All attribute tasks, convolutional layer and full articulamentum are shared, different task is adjusted by multitask weight key-course
Weight so as to carrying out combination learning.
The definition for the task that network needs learn is listed first.
1st, the definition of task.Attribute forecast task is considered as a recurrence task by the present invention.For positive sample, 1 is labeled as,
For negative sample, -1 is labeled as.Assuming that we finally obtain the sharing feature R4096 of 4096 dimensions, then R4096 →
(+1, −1)Carry out the recurrence of each attribute, it is assumed that we have k attribute task, then final R4096 → Rk, pass through connection
Study is closed, returns out all properties task.When prediction, if the attribute that prediction obtains>=threshold value, then it is judged to depositing
In this attribute, if<Threshold value, then it is judged to that this attribute is not present.
Next 2 important composition modules of this deep neural network model are introduced.More including adaptive weighting
The training method of model of being engaged in and the update method of each task respective weights.
2nd, the multi task model training algorithm of adaptive weighting
In the adaptive task weight adjustment model that this method proposes, weight at different moments can in training of each task
Change, the change of weight is to be determined according to each task verifying the error size on collecting and variation tendency.
Algorithm 1 in annex illustrates the specific training process of model.It is the iteration time during model training wherein to remember c
Number, λ is weight vectors, is used to refer to the weight of all tasks, and it is 1 that this is vector initialising, is represented initial when, is owned
The all shared identical weight of task, val_loss_list is the data structure that collection error amount is verified for storing, and note k is appoints
The iterations parameter of weight of being engaged in renewal.The learning process of model is divided into 3 steps:
1) in network training, when iterations c is less than iterations higher limit, the error amount in validation data set is calculated
Val_loss, and store these error amounts using val_loss_list;
2) taken turns per k, we carry out calculating renewal at the respective weights λ value to all tasks, and the algorithm of this renewal is designated as update_
weights();
3) after the weight λ for calculating different task, processing is weighted to the loss values of passback according to these weights, with calculating
Obtained weighted_loss updates parameter in network.
3rd, the more new algorithm of weight
Algorithm 2 in annex specifically presents the weight more new algorithm of model.According to weight adaptive algorithm thinking, the change of weight
Change is to be determined according to each task verifying the error size on collecting and variation tendency, thus more new algorithm can specifically divide
For following 6 steps:
1) according to the data stored in val_loss_list, mean error pre_ of each task in previous stage is calculated
mean。
2) according to the data stored in val_loss_list, mean error cur_ of each task in the current generation is calculated
mean。
3) according to current error cur_mean and the error pre_mean of previous stage, the trend of calculation error change
trend。
4) norm_trend is normalized to the trend of error.
5) norm_loss is normalized to the size of error amount.
6) the weight λ of each task is calculated according to the trend of the size of error amount and error.
Bibliography
1.Zhanpeng Zhang, Ping Luo, Chen Change Loy, and Xiaoou Tang. 2014.
Facial landmark detection by deep multi-task learning. In ECCV.
2.Zhanpeng Zhang, Ping Luo, Chen Change Loy, and Xiaoou Tang. 2016.
Learning deep representation for face alignment with auxiliary attributes.
IEEE transactions on pattern analysis and machine intelligence 38, 5 (2016),
918–930.
3. Abrar H Abdulnabi, Gang Wang, Jiwen Lu, and Kui Jia. 2015. Multi-task
cnn model for attribute prediction. IEEE TMM (2015).
4.Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. 2016. Joint
Face Detection and Alignment Using Multitask Cascaded Convolutional Networks.IEEE Signal Processing Letters 23, 10 (2016), 1499–1503.。
Annex
Algorithm 1:The multi task model training algorithm of adaptive weighting
Algorithm 2:The more new algorithm of weight
。
Claims (3)
1. a kind of multi task model training method of adaptive weighting adjustment, it is characterised in that comprise the following steps that:
(1)The collection of face and pedestrian's picture and the other mark of Attribute class
Collect certain original image data, the training for human body attributive analysis model;There is corresponding people for every pictures
The attribute classification mark of body;Face character includes:Whether wear glasses, if wear masks, if wear sunglasses, if make up, be
Whether no youth, hair length, hair crimp, eyebrows bushed together, eyes size, if be double-edged eyelid, nose whether Gao Ting, if
There is double chin etc.;The attribute of pedestrian includes sleeve length, lower body garment length, garment language, knapsack, handbag, upper body clothes
Color and lower body garment color etc.;
(2)Build deep neural network
Including:Input layer, basic network, multitask weight key-course, basic network include convolutional layer, full articulamentum, pond layer;
Infrastructure network framework uses ResNet-50 frameworks;Wherein:
Input layer is responsible for receiving input;
Input picture passes through the convolutional layer conv1 of first layer, then carries out feature extraction through pond layer, and by full articulamentum;Its
In, pond layer is connected with after first and the convolutional layer of last, pond layer is done to adjacent region thresholding to be polymerize so that network
The certain deformation of tolerable;After convolution and pondization operation, the feature that extraction is obtained inputs full articulamentum, and full articulamentum is
One linear transformation is done to the feature of input, it is pre- so as to complete attribute by the Projection Character of input to a more preferable subspace
Survey task;The last of network is multitask weight key-course, is responsible for calculating the difference value between the attribute and markup information of prediction, and
Adaptive adjustment to weight is completed by backpropagation;
(3)Train deep neural network
Multitask human body attributive analysis model is trained using the multi task model training method of adaptive weighting adjustment, introduces one
The algorithm of the individual weight that can dynamically adjust each task in the training process based on validation error;The weight of each task exists
Training can all change at different moments, and the change of weight is to verify the error size on collecting and change according to each task
Change trend determines, by constantly iterating to calculate and backpropagation, optimizes the parameter in depth network model;
(4)Utilize step(3)In depth network model, carry out human body attribute forecast
After the completion of the training of above-mentioned depth model, for a given face or pedestrian's picture, export in the image for people
The prediction result of face attribute or pedestrian's attribute.
2. the multi task model training method of adaptive weighting adjustment according to claim 1, it is characterised in that except last
Multitask weight key-course, outside full articulamentum and all pond layers, every layer is respectively connected with non-linear layer below in network, adopts
By the use of Relu as activation primitive, function representation is f (x)=max (0, x).
3. the multi task model training method of adaptive weighting adjustment according to claim 1, it is characterised in that the depth
Spend in neural network model, the renewal side of the training method of the multi task model of adaptive weighting and each task respective weights
Method comprises the following steps that:
(1)The multi task model training method of adaptive weighting
In adaptive task weight adjusts model, weight can change at different moments in training of each task, power
The change of weight is to be determined according to each task verifying the error size on collecting and variation tendency;Specifically training process is:
It is the iterations during model training to remember c, and λ is weight vectors, is used to refer to the weight of all tasks, this vector
1 is initialized as, is represented initial when, all shared identical weight of all tasks, val_loss_list is to be used to store
The data structure of checking collection error amount, note k are the iterations parameter of task weight renewal;The learning process of model is divided into 3
Step:
1)In network training, when iterations c is less than iterations higher limit, the error amount in validation data set is calculated
Val_loss, and store these error amounts using val_loss_list;
2)Taken turns per k, calculating renewal is carried out to the respective weights λ value of all tasks, the algorithm of this renewal is designated as update_
weights();
3)After the weight λ for calculating different task, processing is weighted to the loss values of passback according to these weights, with calculating
Obtained weighted_loss updates parameter in network;
(2)The update method of weight
According to weight adaptive algorithm thinking, the change of weight is to verify the error size on collecting and change according to each task
Change trend determines, is specifically divided into following 6 steps:
1)According to the data stored in val_loss_list, mean error pre_mean of each task in previous stage is calculated;
2)According to the data stored in val_loss_list, mean error cur_mean of each task in the current generation is calculated;
3)According to current error cur_mean and the error pre_mean of previous stage, the trend trend of calculation error change;
4)Norm_trend is normalized to the trend of error;
5)Norm_loss is normalized to the size of error amount;
6)The weight λ of each task is calculated according to the trend of the size of error amount and error.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710603212.2A CN107563279B (en) | 2017-07-22 | 2017-07-22 | Model training method for adaptive weight adjustment aiming at human body attribute classification |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710603212.2A CN107563279B (en) | 2017-07-22 | 2017-07-22 | Model training method for adaptive weight adjustment aiming at human body attribute classification |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107563279A true CN107563279A (en) | 2018-01-09 |
CN107563279B CN107563279B (en) | 2020-12-22 |
Family
ID=60973722
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710603212.2A Active CN107563279B (en) | 2017-07-22 | 2017-07-22 | Model training method for adaptive weight adjustment aiming at human body attribute classification |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107563279B (en) |
Cited By (21)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108345912A (en) * | 2018-04-25 | 2018-07-31 | 电子科技大学中山学院 | Commodity rapid settlement system based on RGBD information and deep learning |
CN108648072A (en) * | 2018-05-18 | 2018-10-12 | 深圳灰猫科技有限公司 | Internet finance lending risk evaluating system based on user credit dynamic grading |
CN109145981A (en) * | 2018-08-17 | 2019-01-04 | 上海非夕机器人科技有限公司 | Deep learning automation model training method and equipment |
CN109359562A (en) * | 2018-09-29 | 2019-02-19 | 佳都新太科技股份有限公司 | Target identification method, device, target identification equipment and storage medium |
CN109711343A (en) * | 2018-12-27 | 2019-05-03 | 北京思图场景数据科技服务有限公司 | Behavioral structure method based on the tracking of expression, gesture recognition and expression in the eyes |
CN109712103A (en) * | 2018-11-26 | 2019-05-03 | 深圳艺达文化传媒有限公司 | From the eyes processing method and Related product of the Thunder God picture that shoots the video |
CN110084216A (en) * | 2019-05-06 | 2019-08-02 | 苏州科达科技股份有限公司 | Human face recognition model training and face identification method, system, equipment and medium |
CN110263949A (en) * | 2019-06-21 | 2019-09-20 | 安徽智寰科技有限公司 | Merge the data processing method and system of machine mechanism and intelligent algorithm system |
CN110348416A (en) * | 2019-07-17 | 2019-10-18 | 北方工业大学 | Multi-task face recognition method based on multi-scale feature fusion convolutional neural network |
CN110516512A (en) * | 2018-05-21 | 2019-11-29 | 北京中科奥森数据科技有限公司 | Training method, pedestrian's attribute recognition approach and the device of pedestrian's attributive analysis model |
WO2019233226A1 (en) * | 2018-06-05 | 2019-12-12 | 腾讯科技(深圳)有限公司 | Face recognition method, classification model training method and device, storage medium and computer device |
CN110674756A (en) * | 2019-09-25 | 2020-01-10 | 普联技术有限公司 | Human body attribute recognition model training method, human body attribute recognition method and device |
WO2020078200A1 (en) * | 2018-10-19 | 2020-04-23 | 中兴通讯股份有限公司 | Data processing method and device, and computer-readable storage medium |
TWI709090B (en) * | 2019-08-30 | 2020-11-01 | 阿證科技股份有限公司 | Neural-like artificial intelligence decision network core system and its information processing method |
CN113408439A (en) * | 2021-06-23 | 2021-09-17 | 广东工业大学 | Individual gait recognition method based on trinocular visual data |
CN113673635A (en) * | 2020-05-15 | 2021-11-19 | 复旦大学 | Self-supervision learning task-based hand-drawn sketch understanding deep learning method |
CN114155589A (en) * | 2021-11-30 | 2022-03-08 | 北京百度网讯科技有限公司 | Image processing method, device, equipment and storage medium |
CN115019349A (en) * | 2022-08-09 | 2022-09-06 | 中科视语(北京)科技有限公司 | Image analysis method, image analysis device, electronic equipment and storage medium |
WO2022227772A1 (en) * | 2021-04-27 | 2022-11-03 | 北京百度网讯科技有限公司 | Method and apparatus for training human body attribute detection model, and electronic device and medium |
CN115984804A (en) * | 2023-03-14 | 2023-04-18 | 安徽蔚来智驾科技有限公司 | Detection method based on multi-task detection model and vehicle |
CN117152566A (en) * | 2023-10-30 | 2023-12-01 | 苏州元脑智能科技有限公司 | Classification model training method, model, classification method and product |
Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1925543A (en) * | 2006-08-10 | 2007-03-07 | 威盛电子股份有限公司 | Weight adjusting module and weight adjusting method |
CN101968780A (en) * | 2010-09-28 | 2011-02-09 | 天津大学 | Nonparametric regression method |
CN102722577A (en) * | 2012-06-05 | 2012-10-10 | 中兴通讯股份有限公司 | Method and device for determining dynamic weights of indexes |
CN104866810A (en) * | 2015-04-10 | 2015-08-26 | 北京工业大学 | Face recognition method of deep convolutional neural network |
CN104915730A (en) * | 2015-06-09 | 2015-09-16 | 西北工业大学 | Device multi-attribute maintenance decision method based on weight |
CN105243356A (en) * | 2015-09-10 | 2016-01-13 | 北京大学 | Method of building pedestrian detection model and device and pedestrian detection method |
US20160196480A1 (en) * | 2014-05-05 | 2016-07-07 | Atomwise Inc. | Systems and methods for applying a convolutional network to spatial data |
CN105976207A (en) * | 2016-05-11 | 2016-09-28 | 山东大学 | Information search result generation method and system based on multi-attribute dynamic weight distribution |
CN106575367A (en) * | 2014-08-21 | 2017-04-19 | 北京市商汤科技开发有限公司 | A method and a system for facial landmark detection based on multi-task |
CN106934392A (en) * | 2017-02-28 | 2017-07-07 | 西交利物浦大学 | Vehicle-logo recognition and attribute forecast method based on multi-task learning convolutional neural networks |
-
2017
- 2017-07-22 CN CN201710603212.2A patent/CN107563279B/en active Active
Patent Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1925543A (en) * | 2006-08-10 | 2007-03-07 | 威盛电子股份有限公司 | Weight adjusting module and weight adjusting method |
CN101968780A (en) * | 2010-09-28 | 2011-02-09 | 天津大学 | Nonparametric regression method |
CN102722577A (en) * | 2012-06-05 | 2012-10-10 | 中兴通讯股份有限公司 | Method and device for determining dynamic weights of indexes |
US20160196480A1 (en) * | 2014-05-05 | 2016-07-07 | Atomwise Inc. | Systems and methods for applying a convolutional network to spatial data |
CN106575367A (en) * | 2014-08-21 | 2017-04-19 | 北京市商汤科技开发有限公司 | A method and a system for facial landmark detection based on multi-task |
CN104866810A (en) * | 2015-04-10 | 2015-08-26 | 北京工业大学 | Face recognition method of deep convolutional neural network |
CN104915730A (en) * | 2015-06-09 | 2015-09-16 | 西北工业大学 | Device multi-attribute maintenance decision method based on weight |
CN105243356A (en) * | 2015-09-10 | 2016-01-13 | 北京大学 | Method of building pedestrian detection model and device and pedestrian detection method |
CN105976207A (en) * | 2016-05-11 | 2016-09-28 | 山东大学 | Information search result generation method and system based on multi-attribute dynamic weight distribution |
CN106934392A (en) * | 2017-02-28 | 2017-07-07 | 西交利物浦大学 | Vehicle-logo recognition and attribute forecast method based on multi-task learning convolutional neural networks |
Cited By (32)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108345912A (en) * | 2018-04-25 | 2018-07-31 | 电子科技大学中山学院 | Commodity rapid settlement system based on RGBD information and deep learning |
CN108648072A (en) * | 2018-05-18 | 2018-10-12 | 深圳灰猫科技有限公司 | Internet finance lending risk evaluating system based on user credit dynamic grading |
CN110516512B (en) * | 2018-05-21 | 2023-08-25 | 北京中科奥森数据科技有限公司 | Training method of pedestrian attribute analysis model, pedestrian attribute identification method and device |
CN110516512A (en) * | 2018-05-21 | 2019-11-29 | 北京中科奥森数据科技有限公司 | Training method, pedestrian's attribute recognition approach and the device of pedestrian's attributive analysis model |
WO2019233226A1 (en) * | 2018-06-05 | 2019-12-12 | 腾讯科技(深圳)有限公司 | Face recognition method, classification model training method and device, storage medium and computer device |
US11335124B2 (en) | 2018-06-05 | 2022-05-17 | Tencent Technology (Shenzhen) Company Limited | Face recognition method and apparatus, classification model training method and apparatus, storage medium and computer device |
CN109145981A (en) * | 2018-08-17 | 2019-01-04 | 上海非夕机器人科技有限公司 | Deep learning automation model training method and equipment |
CN109145981B (en) * | 2018-08-17 | 2021-12-07 | 上海非夕机器人科技有限公司 | Deep learning automatic model training method and equipment |
CN109359562A (en) * | 2018-09-29 | 2019-02-19 | 佳都新太科技股份有限公司 | Target identification method, device, target identification equipment and storage medium |
WO2020078200A1 (en) * | 2018-10-19 | 2020-04-23 | 中兴通讯股份有限公司 | Data processing method and device, and computer-readable storage medium |
CN109712103A (en) * | 2018-11-26 | 2019-05-03 | 深圳艺达文化传媒有限公司 | From the eyes processing method and Related product of the Thunder God picture that shoots the video |
CN109712103B (en) * | 2018-11-26 | 2021-07-30 | 温岭卓致智能科技有限公司 | Eye processing method for self-shot video Thor picture and related product |
CN109711343A (en) * | 2018-12-27 | 2019-05-03 | 北京思图场景数据科技服务有限公司 | Behavioral structure method based on the tracking of expression, gesture recognition and expression in the eyes |
CN110084216B (en) * | 2019-05-06 | 2021-11-09 | 苏州科达科技股份有限公司 | Face recognition model training and face recognition method, system, device and medium |
CN110084216A (en) * | 2019-05-06 | 2019-08-02 | 苏州科达科技股份有限公司 | Human face recognition model training and face identification method, system, equipment and medium |
CN110263949B (en) * | 2019-06-21 | 2021-08-31 | 安徽智寰科技有限公司 | Data processing method and system fusing machine mechanism and artificial intelligence algorithm system |
CN110263949A (en) * | 2019-06-21 | 2019-09-20 | 安徽智寰科技有限公司 | Merge the data processing method and system of machine mechanism and intelligent algorithm system |
CN110348416A (en) * | 2019-07-17 | 2019-10-18 | 北方工业大学 | Multi-task face recognition method based on multi-scale feature fusion convolutional neural network |
TWI709090B (en) * | 2019-08-30 | 2020-11-01 | 阿證科技股份有限公司 | Neural-like artificial intelligence decision network core system and its information processing method |
CN110674756A (en) * | 2019-09-25 | 2020-01-10 | 普联技术有限公司 | Human body attribute recognition model training method, human body attribute recognition method and device |
CN110674756B (en) * | 2019-09-25 | 2022-07-05 | 普联技术有限公司 | Human body attribute recognition model training method, human body attribute recognition method and device |
CN113673635B (en) * | 2020-05-15 | 2023-09-01 | 复旦大学 | Hand-drawn sketch understanding deep learning method based on self-supervision learning task |
CN113673635A (en) * | 2020-05-15 | 2021-11-19 | 复旦大学 | Self-supervision learning task-based hand-drawn sketch understanding deep learning method |
WO2022227772A1 (en) * | 2021-04-27 | 2022-11-03 | 北京百度网讯科技有限公司 | Method and apparatus for training human body attribute detection model, and electronic device and medium |
CN113408439A (en) * | 2021-06-23 | 2021-09-17 | 广东工业大学 | Individual gait recognition method based on trinocular visual data |
CN114155589B (en) * | 2021-11-30 | 2023-08-08 | 北京百度网讯科技有限公司 | Image processing method, device, equipment and storage medium |
CN114155589A (en) * | 2021-11-30 | 2022-03-08 | 北京百度网讯科技有限公司 | Image processing method, device, equipment and storage medium |
CN115019349A (en) * | 2022-08-09 | 2022-09-06 | 中科视语(北京)科技有限公司 | Image analysis method, image analysis device, electronic equipment and storage medium |
CN115019349B (en) * | 2022-08-09 | 2022-11-04 | 中科视语(北京)科技有限公司 | Image analysis method, image analysis device, electronic equipment and storage medium |
CN115984804A (en) * | 2023-03-14 | 2023-04-18 | 安徽蔚来智驾科技有限公司 | Detection method based on multi-task detection model and vehicle |
CN115984804B (en) * | 2023-03-14 | 2023-07-07 | 安徽蔚来智驾科技有限公司 | Detection method based on multitasking detection model and vehicle |
CN117152566A (en) * | 2023-10-30 | 2023-12-01 | 苏州元脑智能科技有限公司 | Classification model training method, model, classification method and product |
Also Published As
Publication number | Publication date |
---|---|
CN107563279B (en) | 2020-12-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107563279A (en) | The model training method adjusted for the adaptive weighting of human body attributive classification | |
CN107766850B (en) | Face recognition method based on combination of face attribute information | |
CN108764207B (en) | Face expression recognition method based on multitask convolutional neural network | |
CN110443189B (en) | Face attribute identification method based on multitask multi-label learning convolutional neural network | |
EP3767522A1 (en) | Image recognition method and apparatus, and terminal and storage medium | |
CN106650693B (en) | Multi-feature fusion recognition algorithm for face comparison | |
CN109033938A (en) | A kind of face identification method based on ga s safety degree Fusion Features | |
CN107871100A (en) | The training method and device of faceform, face authentication method and device | |
CN106778796A (en) | Human motion recognition method and system based on hybrid cooperative model training | |
CN109815826A (en) | The generation method and device of face character model | |
CN106096538A (en) | Face identification method based on sequencing neural network model and device | |
CN107194341A (en) | The many convolution neural network fusion face identification methods of Maxout and system | |
CN109344713B (en) | Face recognition method of attitude robust | |
CN107808389A (en) | Unsupervised methods of video segmentation based on deep learning | |
CN105426908B (en) | A kind of substation's attributive classification method based on convolutional neural networks | |
CN106127815A (en) | A kind of tracking merging convolutional neural networks and system | |
CN108446676B (en) | Face image age discrimination method based on ordered coding and multilayer random projection | |
CN110781829A (en) | Light-weight deep learning intelligent business hall face recognition method | |
CN104915658B (en) | A kind of emotion component analyzing method and its system based on emotion Distributed learning | |
CN111339988A (en) | Video face recognition method based on dynamic interval loss function and probability characteristic | |
CN110633624B (en) | Machine vision human body abnormal behavior identification method based on multi-feature fusion | |
CN108446672A (en) | A kind of face alignment method based on the estimation of facial contours from thick to thin | |
CN106228575A (en) | Merge convolutional neural networks and the tracking of Bayesian filter and system | |
CN107704848A (en) | A kind of intensive face alignment method based on multi-constraint condition convolutional neural networks | |
CN110188656A (en) | The generation and recognition methods of multi-orientation Face facial expression 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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |