CN107563279B - Model training method for adaptive weight adjustment aiming at human body attribute classification - Google Patents
Model training method for adaptive weight adjustment aiming at human body attribute classification Download PDFInfo
- Publication number
- CN107563279B CN107563279B CN201710603212.2A CN201710603212A CN107563279B CN 107563279 B CN107563279 B CN 107563279B CN 201710603212 A CN201710603212 A CN 201710603212A CN 107563279 B CN107563279 B CN 107563279B
- Authority
- CN
- China
- Prior art keywords
- task
- weight
- training
- layer
- error
- 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.)
- Active
Links
Images
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention belongs to the technical field of computer vision image processing, and particularly relates to a multi-task model training method for self-adaptive weight adjustment aiming at human body attribute classification. The invention provides a novel multi-task model, and the corresponding weight value of each task is self-adaptively and dynamically adjusted in the training process by introducing an algorithm for updating the corresponding task weight based on the verification error and the change trend. The method comprises the following specific steps: (1) collecting human face and pedestrian pictures and labeling attribute categories; (2) constructing a deep neural network; (3) training a deep neural network; (4) predicting the human body attribute by using a depth network model; the method has the advantages of high speed, high accuracy, good robustness and the like, and is very suitable for practical application of human body related detection, identification, classification and the like.
Description
Technical Field
The invention belongs to the technical field of computer vision image processing, and particularly relates to a multi-task model training method for adaptive weight adjustment aiming at human body attribute classification.
Background
The human body attribute analysis technology is widely applied to real life, such as an environmental safety monitoring system, a traffic control and supervision system and the like. Meanwhile, the human body attribute is also a key characteristic for advanced computer vision tasks such as human body weight recognition, clothing matching and the like, so that the human body attribute analysis technology is concerned more and more by researchers. In non-limiting conditions, human body attribute analysis is still a great challenge due to the influence of changes in human body posture, occlusion, light and the like.
With the revival of the convolutional neural network, the deep multitask network predicts the human body attribute by a method of sharing characteristic representation on different types of attributes, and all the attributes of parameters of other layers are shared except the highest layer aiming at a single attribute task in the network. Such an approach can create two problems: 1. when the task gap between the two attributes is large, insufficient violent conversion can impair the performance of the learner. 2. Such a multi-task learning method would preset and fix the respective weights of each task in the training, and thus does not take into account the differences and associations between different tasks. Although the learning method based on multiple tasks is widely applied to attribute prediction, based on the above two problems, the existing multi-task learning method has certain defects, and because the multiple tasks interfere with each other in the learning process, how to effectively adjust and control the weights of different tasks in the training process cannot provide an excellent solution.
In the multi-task model training, the text [1,2] adopts the task of simultaneously predicting auxiliary attributes such as gender, expression and appearance to improve the detection performance of the target task aiming at the detection task of the face key points. Article [3] proposes a convolutional neural network to learn face attributes that allows some visual information to be shared between different tasks. Document [4] exploits the internal correlation between face detection and alignment to improve the performance of deep cascaded multi-tasking networks. However, the control of the adjustment of the weights for the different tasks in the multitask model remains a difficult problem.
In order to solve the problems to be solved, different from the existing attribute analysis method, the invention introduces an algorithm which is based on verification errors and can dynamically adjust the weight of each task in the training process, thereby providing a multi-task model training method with self-adaptive weight adjustment to train a multi-task human body attribute analysis model.
Disclosure of Invention
The invention aims to provide a multi-task model training method for self-adaptive weight adjustment to train a multi-task human body attribute analysis model.
The problem to be solved urgently is the weight adjustment control aiming at different tasks in the multi-task model. In order to solve the problem, the invention provides a novel multitask model, and an algorithm which is based on verification errors and can dynamically adjust the weight of each task in a training process is introduced into the multitask model, so that the corresponding weight value of each task is adaptively and dynamically adjusted in the training process. Specifically, the generalization ability is adopted to measure the importance of a task, and a higher weight is set for a model with poor generalization ability. For each attribute analysis task, its error and the variation trend of the error on the verification set are analyzed. For a task with a large error, the task is shown to have difficulty; for the task with large error trend variation, the learnability of the task model is better, and higher weight parameters are given. By the weight adjusting strategy, each attribute task can be fully learned without overfitting, and the innovation of the invention is highlighted.
The invention provides a multitask model training method for self-adaptive weight adjustment, which comprises the following specific steps:
(1) collecting human face and pedestrian pictures and labeling attribute categories; the method needs to collect certain original picture data for training a human body attribute analysis model; labeling attribute types of corresponding human bodies for each picture; the attributes of the human face comprise whether glasses are worn or not, whether a mask is worn or not, whether sunglasses are worn or not, whether makeup is made or not, whether the hair is young or not, whether the hair is curled or not, whether the eyebrows are thick or thick, whether the eyes are big or small, whether the nose is tall and straight, whether the two chin exist or not and the like; the attributes of the pedestrian include sleeve length, lower body garment length, garment style, backpack, handbag, upper body garment color, lower body garment color, and the like;
(2) constructing a deep neural network; the basic network structure framework adopts a ResNet-50 framework; FIG. 1 is a structure of a designed deep neural network; the method comprises the following steps: the system comprises an input layer, a basic network and a multi-task weight control layer, wherein the basic network comprises a convolutional layer, a full connection layer, a pooling layer and the like; wherein:
the input layer is responsible for accepting input;
the input image passes through a first layer of convolution layer conv1 and then passes through a pooling layer, and feature extraction is carried out through 16 alternate convolution modules (full connection layers); the first and last convolution layers are connected with a pooling layer, and the pooling layer can aggregate adjacent region values, so that the network can tolerate certain deformation and enhance the translation and rotation invariance of the learning characteristics; after convolution and pooling operation, inputting the extracted features into a full-link layer, wherein the full-link layer performs linear transformation on the input features and can project the input features to a better subspace, so that an attribute prediction task is completed; the last part of the network is a multitask weight control layer which is responsible for calculating the difference value between the predicted attribute and the labeling information and completing the self-adaptive adjustment of the weight through back propagation;
(3) training a deep neural network; training a multi-task human body attribute analysis model by adopting a multi-task model training method with self-adaptive weight adjustment, and introducing an algorithm which is based on verification errors and can dynamically adjust the weight of each task in the training process; the weight of each task changes at different training moments, the change of the weight is determined according to the error magnitude and the change trend of each task on the verification set, and parameters in the deep network model are optimized through continuous iterative computation and back propagation;
(4) performing a human body attribute prediction task by using the deep network model in the step 3; after the training of the depth model is completed, for a given face or pedestrian picture, a prediction result of the face attribute or the pedestrian attribute in the image can be output.
The innovation of the invention is that:
1. a multi-task model training method with self-adaptive weight adjustment is provided for training a multi-task human body attribute analysis model. The weight parameters of each task are updated autonomously according to the generalization ability of the corresponding learner of each task, and the problem of weight adjustment control for different tasks in the multi-task model is solved in a breakthrough manner;
2. and adding a multi-task weight control layer to the deep neural network for performing the human body attribute analysis task, and introducing an algorithm which is based on verification errors and can dynamically adjust the weight of each task in the training process. In the end-to-end training process, the corresponding weight parameters of each task are adjusted and controlled by using a back propagation mode, so that the generalization capability of the model is enhanced. By adopting the weight updating training method, the accuracy of the analysis and prediction tasks of the human face attributes and the human body attributes is obviously improved.
Drawings
FIG. 1 is a schematic diagram of a network model framework for human body attribute analysis.
Detailed Description
Step 1, collecting human face and pedestrian pictures and labeling attribute categories. The categories of the attributes of the human face comprise external attributes of the human face and internal attributes of the human face, and the external attributes of the human face comprise whether glasses are worn, whether a mask is worn, whether sunglasses are worn, whether makeup is performed and the like. The intrinsic attributes of the face can be subdivided into global attributes and local attributes. The overall attributes include gender, youth, face value, and the like. Local attributes focus on the details of five sense organs of the human face, including hair length, whether hair curls, eyebrow thickness, eye size, whether double eyelids are present, whether nose is tall and straight, whether double chin is present, and the like. Pedestrian attributes include sleeve length, lower body garment length, style of garment, backpack, handbag, upper body garment color, lower body garment color, and the like.
And 2, constructing a deep neural network. The basic network structure framework adopts a ResNet-50 framework, and FIG. 1 shows the structure of a designed deep neural network. Except the last multitask weight control layer, the full connection layer and all the pooling layers, each layer in the network is connected with a nonlinear layer, Relu is adopted as an activation function, and the function is expressed as f (x) = max (0, x).
The input image first passes through the first layer of convolution layer conv1, then passes through the pooling layer, and is feature extracted by 16 alternating convolution modules. The first and last convolution layers are connected with a pooling layer, and the pooling layer can aggregate adjacent region values, so that the network can tolerate certain deformation and enhance the translation and rotation invariance of the learning characteristics. After convolution and pooling operation, the extracted features are input into a full-link layer, the full-link layer performs linear transformation on the input features, and the input features can be projected to a better subspace, so that an attribute prediction task is completed. And the last part of the network is a multitask weight control layer which is responsible for calculating the difference value between the predicted attribute and the labeling information.
In general, the input layer is responsible for accepting input. And performing feature extraction on the picture through the combination of the alternate convolutional layers, the nonlinear layers and the pooling layers. The fully connected layer can map the obtained features. And the last multitask weight control layer is responsible for calculating the prediction error of the network and finishing the self-adaptive adjustment of the weight through back propagation.
And 3, training the deep neural network. And after the human face, the human body picture and the label information of the corresponding attribute are prepared, training of the deep network is carried out. The multi-task model training method for adaptive weight adjustment proposed by the present invention is described in detail with reference to fig. 1. After an input training image passes through a plurality of convolution layers and pooling layers, inputting the extracted feature graph into a multi-task weight control layer for calculation, wherein the weight of each task changes at different moments of training, and the change of the weight is determined according to the error magnitude and the change trend of each task on a verification set. The information of the network will be described in detail below.
Fig. 1 shows the main architecture of the method. There are 3 main components, convolutional layers, fully-connected layers, and multi-tasking weight control layers. All attribute tasks share the convolution layer and the full connection layer, and the weights of different tasks are adjusted through the multi-task weight control layer so as to perform joint learning.
The definition of the tasks that the network needs to learn is listed first.
1. And (4) defining the task. The invention regards the attribute prediction task as a regression task. For positive samples, labeled 1, and for negative samples, labeled-1. Assuming that we finally get 4096-dimensional shared feature R4096, then R4096 → (+ 1, -1) performs regression of each attribute, assuming we have k attribute tasks, then finally R4096 → Rk, through joint learning, regresses all attribute tasks. At the time of prediction, if the predicted attribute > = threshold, it is determined that this attribute exists, and if < threshold, it is determined that this attribute does not exist.
Next, 2 important building blocks of this deep neural network model are described. The method comprises a training method of a multi-task model with self-adaptive weights and an updating method of the corresponding weights of each task.
2. Adaptive weight multi-task model training algorithm
In the self-adaptive task weight adjustment model provided by the method, the weight of each task changes at different training moments, and the change of the weight is determined according to the error magnitude and the change trend of each task on the verification set.
Algorithm 1 in the appendix shows the specific training process of the model. Wherein c is the iteration number in the model training process, λ is a weight vector for indicating the weights of all tasks, the vector is initialized to 1, which means that all tasks share the same weight at the time of initialization, val _ loss _ list is a data structure for storing an error value of a verification set, and k is an iteration number parameter for updating the task weights. The learning process of the model is divided into 3 steps:
1) in network training, when the iteration number c is smaller than the iteration number upper limit value, calculating an error value val _ loss on the verification data set, and storing the error values by using the val _ loss _ list;
2) every k rounds, calculating and updating corresponding weight lambda values of all tasks, and recording an updated algorithm as update _ weights ();
3) after weights lambda of different tasks are calculated, weighting processing is carried out on the returned loss value according to the weights, and the parameters in the network are updated by the weighted _ loss obtained through calculation.
3. Weight updating algorithm
1) the average error pre mean of each task at the previous stage is calculated from the data stored in val _ loss _ list.
2) And calculating the average error cur _ mean of each task at the current stage according to the data stored in val _ loss _ list.
3) The trend of the error change trend is calculated from the current error cur mean and the error pre mean of the previous stage.
4) The trend of the error is normalized norm _ trend.
5) The magnitude of the error value is normalized norm _ loss.
6) And calculating the weight lambda of each task according to the magnitude of the error value and the trend of the error.
Reference to the literature
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.。
Appendix
Algorithm 1: adaptive weight multi-task model training algorithm
And 2, algorithm: weight updating algorithm
Claims (1)
1. A multi-task model training method for self-adaptive weight adjustment is characterized by comprising the following specific steps:
(1) collection of face and pedestrian pictures and labeling of attribute categories
Collecting certain original picture data for training a human body attribute analysis model; for each picture, corresponding attribute category labels of the human body are marked; the attributes of the human face include: whether glasses are worn, whether a mask is worn, whether sunglasses are worn, whether makeup is made, whether the eyebrows are young, the hair length, whether the hairs are curled, the eyebrows are thick, the eyes are big and small, whether the eyebrows are double eyelids, whether the nose is high and straight, and whether the chin is double; the attributes of the pedestrian include sleeve length, lower body garment length, garment style, backpack, handbag, upper body garment color and lower body garment color;
(2) constructing a deep neural network
The method comprises the following steps: the system comprises an input layer, a basic network and a multi-task weight control layer, wherein the basic network comprises a convolution layer, a full connection layer and a pooling layer; the basic network structure framework adopts a ResNet-50 framework; wherein:
the input layer is responsible for accepting input;
the input image passes through the convolution layer conv1 of the first layer, then passes through the pooling layer and is subjected to feature extraction through the full-connection layer; wherein, the first and last convolution layers are connected with a pooling layer, and the pooling layer aggregates adjacent region values, so that the network can tolerate certain deformation; after convolution and pooling operation, inputting the extracted features into a full-connection layer, wherein the full-connection layer performs linear transformation on the input features and projects the input features to a better subspace, so that an attribute prediction task is completed; the last part of the network is a multitask weight control layer which is responsible for calculating the difference value between the predicted attribute and the labeling information and completing the self-adaptive adjustment of the weight through back propagation;
(3) training deep neural networks
Training a multi-task human body attribute analysis model by adopting a multi-task model training method with self-adaptive weight adjustment, and introducing an algorithm which is based on verification errors and can dynamically adjust the weight of each task in the training process; the weight of each task changes at different training moments, the change of the weight is determined according to the error magnitude and the change trend of each task on the verification set, and parameters in the deep network model are optimized through continuous iterative computation and back propagation;
(4) predicting the human body attribute by using the deep network model in the step (3)
After the training of the depth model is finished, for a given face or pedestrian picture, outputting a prediction result of the face attribute or the pedestrian attribute in the picture;
except the last multitask weight control layer, the full connection layer and all the pooling layers, a nonlinear layer is connected behind each layer in the network, Relu is adopted as an activation function, and the function is expressed as f (x) = max (0, x);
in the deep neural network model, the training method of the multi-task model with the self-adaptive weight and the updating method of the corresponding weight of each task specifically comprise the following steps:
(1) adaptive weight multi-task model training method
In the self-adaptive task weight adjustment model, the weight of each task changes at different training moments, and the change of the weight is determined according to the error magnitude and the change trend of each task on the verification set; the specific training process is as follows:
c is recorded as iteration times in the model training process, lambda is a weight vector used for indicating the weights of all tasks, the vector is initialized to be 1, which indicates that all tasks share the same weight in the initialization process, val _ loss _ list is a data structure used for storing an error value of a verification set, and k is recorded as an iteration time parameter for updating the task weights; the learning process of the model is divided into 3 steps:
1) in network training, when the iteration number c is smaller than the iteration number upper limit value, calculating an error value val _ loss on the verification data set, and storing the error values by using the val _ loss _ list;
2) every k rounds, calculating and updating corresponding weight lambda values of all tasks, and recording an updated algorithm as update _ weights ();
3) after weights lambda of different tasks are calculated, weighting processing is carried out on the returned loss value according to the weights, and the parameters in the network are updated by the calculated weighted _ loss;
(2) weight updating method
According to the thought of the weight adaptive algorithm, the change of the weight is determined according to the error magnitude and the change trend of each task on the verification set, and the method specifically comprises the following 6 steps:
1) calculating the average error pre _ mean of each task in the previous stage according to the data stored in val _ loss _ list;
2) calculating the average error cur _ mean of each task at the current stage according to the data stored in val _ loss _ list;
3) calculating the trend of the error change according to the current error cur _ mean and the error pre _ mean of the previous stage;
4) normalizing norm _ trend on the trend of the error;
5) normalizing norm _ loss to the magnitude of the error value;
6) and calculating the weight lambda of each task according to the magnitude of the error value and the trend of the 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 CN107563279A (en) | 2018-01-09 |
| CN107563279B true 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) |
Families Citing this family (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 |
| 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 |
| CN108846340B (en) * | 2018-06-05 | 2023-07-25 | 腾讯科技（深圳）有限公司 | Face recognition method and device, classification model training method and device, storage medium and computer 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 |
| CN111079890A (en) * | 2018-10-19 | 2020-04-28 | 中兴通讯股份有限公司 | Data processing method and device and computer readable storage medium |
| 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 |
| 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 |
| 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 |
| 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 |
| CN113177469B (en) * | 2021-04-27 | 2024-04-12 | 北京百度网讯科技有限公司 | Training method and device of human attribute detection model, electronic equipment 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 |
| 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 |
| 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 |
Citations (9)
| 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 |
| 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 |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9373059B1 (en) * | 2014-05-05 | 2016-06-21 | Atomwise Inc. | Systems and methods for applying a convolutional network to spatial data |
-
2017
- 2017-07-22 CN CN201710603212.2A patent/CN107563279B/en active Active
Patent Citations (9)
| 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 |
| 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 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107563279A (en) | 2018-01-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107563279B (en) | Model training method for adaptive weight adjustment aiming at human body attribute classification | |
| US11393206B2 (en) | Image recognition method and apparatus, terminal, and storage medium | |
| CN107766850B (en) | Face recognition method based on combination of face attribute information | |
| Xu et al. | Exploiting low-rank structure from latent domains for domain generalization | |
| CN109033938A (en) | A kind of face identification method based on ga s safety degree Fusion Features | |
| CN115345278A (en) | Structural learning of convolutional neural networks | |
| CN104463191A (en) | Robot visual processing method based on attention mechanism | |
| Li et al. | Model-based online learning with kernels | |
| Zhou et al. | Online support vector machine: A survey | |
| KR20210053052A (en) | Color restoration method and apparatus | |
| Wang et al. | A face-recognition approach using deep reinforcement learning approach for user authentication | |
| Mafeni Mase et al. | Benchmarking deep learning models for driver distraction detection | |
| Angelov et al. | Toward anthropomorphic machine learning | |
| Su et al. | HDL: Hierarchical deep learning model based human activity recognition using smartphone sensors | |
| CN113673510A (en) | Target detection algorithm combining feature point and anchor frame joint prediction and regression | |
| CN112069916B (en) | Face beauty prediction method, device and system and readable storage medium | |
| Tallec et al. | Multi-order networks for action unit detection | |
| CN111401116A (en) | Bimodal emotion recognition method based on enhanced convolution and space-time L STM network | |
| Li et al. | Class-incremental gesture recognition learning with out-of-distribution detection | |
| Zheng et al. | Action recognition based on the modified twostream CNN | |
| CN109002808A (en) | A kind of Human bodys' response method and system | |
| Wu et al. | Self-learning and explainable deep learning network toward the security of artificial intelligence of things | |
| Yu et al. | Evolve: Enhancing unsupervised continual learning with multiple experts | |
| Yao et al. | Regional attention reinforcement learning for rapid object detection | |
| Jain et al. | Real-time eyeglass detection using transfer learning for non-standard facial data. |
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 |