CN110309810A - A kind of pedestrian's recognition methods again based on batch center similarity - Google Patents
A kind of pedestrian's recognition methods again based on batch center similarity Download PDFInfo
- Publication number
- CN110309810A CN110309810A CN201910617855.1A CN201910617855A CN110309810A CN 110309810 A CN110309810 A CN 110309810A CN 201910617855 A CN201910617855 A CN 201910617855A CN 110309810 A CN110309810 A CN 110309810A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- vector
- batch
- recognition methods
- sample
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of pedestrian's recognition methods again based on batch center similarity, comprising: (1) based on classical infrastructure network, removes the last full articulamentum of network, and add additional convolutional layer to establish full convolutional neural networks model;(2) P pedestrian is randomly selected from original training data concentration, each pedestrian randomly selects K image;(3) P*K image for utilizing (2) to obtain is sent into network and is trained to obtain P*K feature vector;(4) center vector is asked to the K feature vector of each pedestrian, obtains P center vector;(5) the corresponding center vector of each of P*K feature vector and non-similar center vector are constituted into a triple and carries out regression optimization.The method of the present invention is simple and easy, has a wide range of application, and can be effectively solved the problems such as misplacing, block in pedestrian's weight identification mission.
Description
Technical field
The invention belongs to technical field of computer vision, more particularly, to a kind of row based on batch center similarity
People's recognition methods again.
Background technique
Pedestrian identify again in people's daily life there are many aspect application, as video monitoring, personage positioning with
Track etc..Pedestrian identifies that (Person Re-identification) refers to again and carries out in the multi-cam network of the non-overlap ken
Pedestrian's matching, i.e., be matched to interested people in multiple scenes, method proposed by the invention seeks to solve this to ask
Topic.With the explosive growth of video data, the cost that target pedestrian is searched and positioned using the mode manually supervised will be more
On the crucial opportunity that higher and inefficient artificial based manner easilys lead to miss cracking of cases, to cause people's life wealth
The heavy losses of production.So accurately and efficiently positioning pedestrian using machine becomes a kind of urgent demand.
Research emphasis is focused in two wherein the most key problems at present: feature extraction and metric learning, wherein measuring
The method of study is mainly based upon cross entropy loss function, triple loss function, spherical surface loss function etc., these loss functions
Universal problem is big with regard to training sample selection difficulty, can not optimize between class distance and inter- object distance etc. simultaneously, therefore generally imitate
Fruit is not ideal enough.
Summary of the invention
In view of the problems of the existing technology, it is easy to need to design a kind of training sample selection, is furthermore possible to optimize simultaneously
The loss function of between class distance and inter- object distance obtains more preferably training effect with this.And the present invention aiming at measuring before
The drawbacks of method, proposes a kind of method of efficient metric learning.The purpose of the present invention is to provide one kind to be based on batch center
The pedestrian of similarity recognition methods again, this method are realized simply, are chosen training sample and are easy, can optimize sample class spacing simultaneously
From and inter- object distance, available efficient measurement effect.
To achieve the above object, according to one aspect of the present invention, a kind of row based on batch center similarity is provided
People's recognition methods again, includes the following steps:
(1) samples selection:
(1.1) training sample is selected from the picture of the collected multiple pedestrian's different angles of multiple cameras, and the same person exists
Morphological differences is bigger under different cameras, wherein each image only includes a pedestrian, P is randomly choosed from training sample
A pedestrian randomly chooses K sample from each pedestrian, may be constructed the batch sample of a P*K in this way, the P and K are pre-
If value;
(1.2) in view of the sample number of each pedestrian is not quite similar, it is possible to which the sample number of part pedestrian is less than selected K
Value, therefore using the method for putting back to formula sampling: i.e. when each pedestrian sample number is less than K, it can repeat to select, until reaching institute
If K value;
(2) data enhance:
(2.1) area that area represents original image is set, s ∈ [A*area, B*area] is taken to represent the area randomly selected
Then size blocks the image section that the area randomly selected is s;Wherein, wherein 0 < A < B < 1, the embodiment of the present invention
Middle selection A=0.02, B=0.4;
(2.2) for the image obtained after blocking in (2.1), zoom to the first fixed size first, then again with
Machine intercepts the image of the second fixed size, and first fixed size and the second fixed size are preset value, and described second is fixed
Size is less than first fixed size;Such as [288,144] size is zoomed to first, then image is intercepted at random
The image of [256,128] size, as training image;
(3) feature extraction:
(3.1) select ResNet50 as supporting network, wherein various pieces all use ReLU function as activation primitive,
Its function definition isThe network has carried out pre-training on ImageNet data set;
(3.2) in the 4th block of the network that (3.1) are selected, the step-length of convolution is reduced to 1 from original 2, in this way
The resolution ratio for the feature vector that can be improved, the feature vector dimension exported at this time are 2048 dimensions;It is original after this layer
The pool layer of ResNet50 is replaced with max_pool for average_pool in order to learn to more distinguishable feature
Then layer increases a full articulamentum, so that output dimension is reduced to 1024 dimensions from 2048 dimensions, the output of this layer is for table
The feature vector of image is levied, F1 is labeled as;After this layer, increase a full articulamentum, output dimension is the class of data set
Not Shuo, such as Market1501 data set, output dimension be just set as 751, for obtained vector, be labeled as F2;
(4) trunk optimization aim:
(4.1) according to step (3.2), having obtained two sizes is P*K, dimension be respectively 1024 and labels (sample
Classification number) feature vector, it may be assumed that F1 and F2.To F1, the mean vector of everyone K feature vector is sought respectively, is labeled as
Ci, i ∈ [1, P], P center vector available in this way;
(4.2) calculation formula of the triple loss function based on batch center of final design are as follows:Wherein TBCL is the contracting of Triplet Batch Center Loss
It writing, S represents similarity formula,M represents the threshold value of setting.In order to simplify calculation amount while be conducive to gradient
Backpropagation, unitization operation is carried out to all feature vector, it may be assumed thatIn turnBy formula as can be seen that purpose be so that each sample, the phase with its central point
It is increasing while smaller and smaller with the similarity of its non-central point like spending, the between class distance of sample can be made more next in this way
Bigger, inter- object distance is smaller and smaller;
(5) optimization aim is assisted:
(5.1) in many tasks it is verified that, the method based on multi-task learning can often be obtained than single task
Better result.According to (4.2), the feature vector F2 that dimension is [P*K, labels] has been obtained, has been carried out using entropy function is intersected
Regression optimization:Wherein yjTrue label is represented,It in this way can further degree of being promoted
Measure the effect of study.
Contemplated above technical scheme through the invention, compared with prior art, the present invention has following technical effect that
(1) method is realized simple: the method for the present invention is compared with the method for previous metric learning, by calculating batch center
Method, clear thinking is simple and effective;
(2) choose sample to be easy: traditional triple loss function needs to choose training sample, such as selection and mesh meticulously
The farthest positive sample of standard specimen sheet and nearest negative sample, the method for the present invention when choosing sample, without the concern for target sample and
The relationship of positive sample, therefore choose sample and be comparatively easy;
(3) strong robustness: the method for traditional some metric learnings, such as center loss function, spherical surface loss function are past
Inter- object distance can not be optimized well toward between class distance can only be optimized, the metric function that the present invention realizes can optimize class simultaneously
Between distance and inter- object distance, therefore performance more Shandong nation;
Detailed description of the invention
Fig. 1 is that the present invention is based on the flow charts of the pedestrian of batch center similarity again recognition methods;
Fig. 2 is the schematic network structure of convolutional neural networks model in the embodiment of the present invention;
Fig. 3 is that center vector generates schematic diagram in the embodiment of the present invention;
Fig. 4 is the measurement effect and other methods contrast schematic diagram in the embodiment of the present invention in test sample.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.As long as in addition, technical characteristic involved in the various embodiments of the present invention described below
Not constituting a conflict with each other can be combined with each other.
Just technical term of the invention is explained and illustrated first below:
Market1501 data set: Market-1501 data set acquires in Tsinghua Campus, summer shooting,
2015 building and openly.It includes being taken by 6 cameras (wherein 5 high-definition cameras and 1 low clear camera)
1501 pedestrians, 32668 pedestrian's rectangle frames detected.Each pedestrian is at least captured by 2 cameras, and at one
There may be multiple images in camera.Training set has 751 people, includes 12936 images, and average everyone has 17.2 training
Data;Test set has 750 people, includes 19732 images, and average everyone has 26.3 test datas.3368 query images
Pedestrian detection rectangle frame manually draw, and the pedestrian detection rectangle frame in gallery is then using deformable component mould
Type (Deformable Part Model, DPM) detector detects.The training set for the fixed quantity that the data set provides
It can be used under single-shot or multi-shot test setting with test set.
CUHK03 data set: CUHK03 is that first extensive pedestrian for being sufficient for deep learning identifies data set again,
The Image Acquisition of the data set is in the campus Hong Kong Chinese University (CUHK).Data contain 1467 different personages, are imaged by 5 Duis
Head acquisition.The data set provides machine detection and manual inspection two datasets, and wherein detection data collection is missed comprising some detections
Difference, closer to actual conditions, average everyone has 9.6 training datas.
DukeMTMC-reID:DukeMTMC data set is the multiple target multiple-camera pedestrian tracking marked on a large scale
Data set.It provides the novel large HD video data set recorded by 8 synchronization cameras, has more than 7000 lists
Video camera track and the independent personage more than 2700, DukeMTMC-reID is that the pedestrian of DukeMTMC data set identifies son again
Collection, and provide the bounding box manually marked.
First matching rate rank-1 and mAP (mean Average Precision).First matching rate is to reflect to identify again
Model performance is the most intuitively also the most key index.Wherein m is the total figure the piece number for inquiring library, Si
It indicates for i-th query image, if successful match, if success Si=1, on the contrary Si=0.In general, identified in pedestrian again
In task, when superiority-inferiority between two models of needs assessment, this index can be compared first, rank-1 higher model is often
Performance is better.This index of mAP considers recall rate (recall) and accuracy (precision) simultaneously, can more integrate
And the objectively performance of evaluation model.In pedestrian's weight identification mission, by the retrieval accuracy of a query sample
(precision) be defined as matching ranking before k image in, ratio shared by correct matching number, it may be assumed thatWherein kcFor
Match correct picture number.And the retrieval recall rate (recall) of the query sample is defined as its image for matching k before ranking
In, correct matched picture number accounts for the ratio of all pedestrian image numbers in candidate library, it may be assumed thatWherein kallTable
Show the picture number for belonging to the same pedestrian with query image in candidate library, is a definite value.When selecting different k, P and
R can occur to change accordingly.K is gradually increased, by corresponding one curve of P and R value paintings, becomes PR figure, calculates some
Area under the PR figure of query sample, referred to as AP value, all query sample AP are worth mean value then to become mAP.
As shown in Figure 1, the present invention is based on the pedestrian of batch center similarity again recognition methods the following steps are included:
(1) samples selection:
(1.1) P pedestrian is randomly choosed from training sample, the K sample of random selection from each pedestrian in this way can be with
Constitute the batch sample of a P*K;
(1.2) in view of the sample number of each pedestrian is not quite similar, it is possible to which the sample number of part pedestrian is less than selected K
Value, therefore using the method for putting back to formula sampling: i.e. when each pedestrian sample number is less than K, it can repeat to select, until reaching institute
If K value;
(2) data enhance:
(2.1) area that area represents original image is set, s ∈ [0.02*area, 0.4*area] representative is taken to randomly select
Then size blocks the image section that the area randomly selected is s;
(2.2) image obtained after blocking at random in (2.1) zooms to [288,144] size, then carries out to image
The image of random interception [256,128] size, as training image;
(3) feature extraction:
(3.1) as shown in Fig. 2, selecting ResNet50 as supporting network, wherein various pieces all use ReLU function to make
For activation primitive, function definition isThe network carries out on ImageNet data set
Pre-training;
(3.2) as shown in Fig. 2, in the 4th block of the network that (3.1) are selected, by the step-length of convolution from original 2
It is reduced to 1, the resolution ratio for the feature vector that can be improved in this way, the feature vector dimension exported at this time is 2048 dimensions;This this
After one layer, the pool layer of original ResNet is that average_pool is replaced with to learn to more distinguishable feature
Max_pool layers, then increase a full articulamentum, so that output dimension is reduced to 1024 dimensions from 2048 dimensions, the output of this layer is just
It is the feature vector for characterizing image, is labeled as F1;After this layer, increase a full articulamentum, output dimension is number
According to the classification number of collection, such as Market1501 data set, output dimension be just set as 751, for obtained vector, be labeled as
F2;
(4) trunk optimization aim:
(4.1) according to step (3.2), having obtained two sizes is P*K, dimension be respectively 1024 and labels (sample
Classification number) feature vector, it may be assumed that F1 and F2.As shown in figure 3, F1 is asked respectively the mean value of everyone K feature vector to
Amount is labeled as Ci, i ∈ [1, P], P center vector available in this way;
(4.2) calculation formula of the triple loss function based on batch center of final design are as follows:Wherein TBCL is the contracting of Triplet Batch Center Loss
It writing, S represents similarity formula,M represents the threshold value of setting.In order to simplify calculation amount while be conducive to gradient
Backpropagation, unitization operation is carried out to all feature vector, it may be assumed thatIn turnBy formula as can be seen that purpose be so that each sample, the phase with its central point
It is increasing while smaller and smaller with the similarity of its non-central point like spending, the between class distance of sample can be made more next in this way
Bigger, inter- object distance is smaller and smaller;
(5) optimization aim is assisted:
(5.1) in many tasks it is verified that, the method based on multi-task learning can often be obtained than single task
Better result.According to (4.2), the feature vector F2 that dimension is [P*K, labels] has been obtained, using intersection entropy function:Wherein yjTrue label is represented,The effect of metric learning can be further promoted in this way
Fruit.
Prove that effectiveness of the invention, the results show present invention can be improved pedestrian's weight below by way of experiment embodiment
The recognition accuracy of identification.
The present invention is existing with representative with 15 on Market1501, CUHK03, DukeMTMC-reID data set
The pedestrian of property knows method for distinguishing again and is compared, table 1 be the method for the present invention and for comparing 15 in control methods at three kinds
The performance of mAP and rank-1 index on public data collection, result value is bigger, and expression performance is better, can see from table, this
Inventive method (i.e. the dated TBCL of table 1) is promoted clearly.
Table 1: distinct methods rank-1 and mAP index on Market1501, CUHK03 and DukeMTMC-reID data set
Performance ,-indicate to lack the data, * indicates reproduction data
10 pedestrians are wherein chosen, each pedestrian has chosen 10 pictures, amounts to 100 pictures, obtains using the present invention
100 feature vectors clustered, according to Fig.4, of a sort sample, the bigger representative of circle are represented in a circle
Clustering Effect is poorer, otherwise better, from fig. 4, it can be seen that Clustering Effect of the invention is best.
As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, not to
The limitation present invention, any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should all include
Within protection scope of the present invention.
Claims (10)
1. a kind of pedestrian's recognition methods again based on batch center similarity, which is characterized in that the method includes the following steps:
(1) samples selection: for training sample, randomly choosing P pedestrian, to each pedestrian, randomly chooses K sample, composition
One batch includes P*K sample, wherein the P and K is preset value, and the number of P is less than pedestrian's total number in sample;
(2) data enhance:
(2.1) for the original image in sample, the part of a random size is blocked using the method blocked at random;
(2.2) for the image obtained after blocking in (2.1), the first fixed size is zoomed to first, it is then random again to cut
The image of the second fixed size is taken, first fixed size and the second fixed size are preset value, second fixed size
Less than first fixed size;
(3) feature extraction:
(3.1) based on classical infrastructure network, remove the last full articulamentum of network, and add additional layer to establish convolution
Neural network model;
(3.2) each batch for intercepting the image of the second fixed size in step (2.2) at random is sent into (3.1) and is set
In network, feature vector is extracted;
(4) trunk optimization aim:
(4.1) it by P*K feature vector obtained in (3.2), is divided according to different people, chooses the K vector of the same person, ask
The average vector of K vector is as a center vector, P central point available in this way;
(4.2) for each feature vector in P*K, its of a sort central point and an inhomogeneous central point are chosen,
In inhomogeneous central point be selected as away from a nearest central point, constitute a triple, utilize triple lose letter
Number carries out regression optimization;
(5) optimization aim is assisted:
(5.1) P*K feature vector obtained in (3.2) is obtained by newly-increased network layer again the feature of tag size to
Amount;
(5.2) by feature vector obtained in (5.1) and its corresponding label, regression optimization is carried out using entropy function is intersected.
2. pedestrian's recognition methods again according to claim 1 based on batch center similarity, which is characterized in that the step
Suddenly (3.1) specifically: select ResNet50 as supporting network, wherein various pieces all use ReLU function as activation letter
Number, function definition are In the 4th block of ResNet50 network, by convolution
Step-length is reduced to 1 from original 2, to improve the resolution ratio of obtained feature vector;After this layer, by original ResNet50's
Pool layers of average_pool replace with max_pool layers, and increase a full articulamentum, and the output of this layer is used to phenogram
The feature vector of picture is labeled as F1;After this layer, increasing a full articulamentum, output dimension is the classification number of data set,
F2 is labeled as to obtained vector.
3. pedestrian's recognition methods again according to claim 1 or 2 based on batch center similarity, which is characterized in that institute
State step (4.1) specifically: according to step (3.2), obtained feature vector F1 and F2, to F1, asked everyone K special respectively
The mean vector of vector is levied, C is labeled asi, i ∈ [1, P], to obtain P center vector.
4. pedestrian's recognition methods again according to claim 1 or 2 based on batch center similarity, which is characterized in that institute
State step (4.2) specifically:
The calculation formula of the triple loss function based on batch center of final design are as follows:Wherein S represents similarity formula,m
Represent the threshold value of setting.
5. pedestrian's recognition methods again according to claim 4 based on batch center similarity, which is characterized in that for letter
Change the backpropagation that calculation amount is conducive to gradient simultaneously, unitization operation carried out to all feature vectors, it may be assumed thatIn turn
6. pedestrian's recognition methods again according to claim 1 or 2 based on batch center similarity, which is characterized in that institute
State step (5.2) specifically: according to the feature vector F2 for the label dimension that (5.1) obtain, returned using entropy function is intersected
Optimization:Wherein yjTrue label is represented,
7. pedestrian's recognition methods again according to claim 1 or 2 based on batch center similarity, which is characterized in that institute
State step (2.1) specifically: set the area that area represents original image, s ∈ [A*area, B*area] representative is taken to randomly select
Then size blocks the image section that the area randomly selected is s, wherein 0 < A < B < 1.
8. pedestrian's recognition methods again according to claim 1 or 2 based on batch center similarity, which is characterized in that institute
State step (2.2) specifically: the image obtained after blocking at random in (2.1) zooms to [288,144] size, then to figure
Image as being intercepted [256,128] size at random.
9. pedestrian's recognition methods again according to claim 1 or 2 based on batch center similarity, which is characterized in that
It is to the pedestrian using formula sample mode is put back to, i.e., heavy if the sample number of certain pedestrian is less than selected K value when selecting K sample
Final election is selected until reaching set K value.
10. pedestrian's recognition methods again according to claim 7 based on batch center similarity, which is characterized in that A=
0.02, B=0.4.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910617855.1A CN110309810B (en) | 2019-07-10 | 2019-07-10 | Pedestrian re-identification method based on batch center similarity |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910617855.1A CN110309810B (en) | 2019-07-10 | 2019-07-10 | Pedestrian re-identification method based on batch center similarity |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110309810A true CN110309810A (en) | 2019-10-08 |
CN110309810B CN110309810B (en) | 2021-08-17 |
Family
ID=68079809
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910617855.1A Active CN110309810B (en) | 2019-07-10 | 2019-07-10 | Pedestrian re-identification method based on batch center similarity |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110309810B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110968735A (en) * | 2019-11-25 | 2020-04-07 | 中国矿业大学 | Unsupervised pedestrian re-identification method based on spherical similarity hierarchical clustering |
CN111091548A (en) * | 2019-12-12 | 2020-05-01 | 哈尔滨市科佳通用机电股份有限公司 | Railway wagon adapter dislocation fault image identification method and system based on deep learning |
CN111666823A (en) * | 2020-05-14 | 2020-09-15 | 武汉大学 | Pedestrian re-identification method based on individual walking motion space-time law collaborative identification |
CN112818755A (en) * | 2021-01-13 | 2021-05-18 | 华中科技大学 | Gait recognition method based on active learning |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR101692747B1 (en) * | 2015-09-09 | 2017-01-04 | 인천대학교 산학협력단 | Improved directional weighted interpolation method for single-sensor camera imaging |
CN107451545A (en) * | 2017-07-15 | 2017-12-08 | 西安电子科技大学 | The face identification method of Non-negative Matrix Factorization is differentiated based on multichannel under soft label |
CN107832711A (en) * | 2017-11-13 | 2018-03-23 | 常州大学 | A kind of recognition methods again of the pedestrian based on transfer learning |
CN108230322A (en) * | 2018-01-28 | 2018-06-29 | 浙江大学 | A kind of eyeground feature detection device based on weak sample labeling |
CN109344787A (en) * | 2018-10-15 | 2019-02-15 | 浙江工业大学 | A kind of specific objective tracking identified again based on recognition of face and pedestrian |
-
2019
- 2019-07-10 CN CN201910617855.1A patent/CN110309810B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR101692747B1 (en) * | 2015-09-09 | 2017-01-04 | 인천대학교 산학협력단 | Improved directional weighted interpolation method for single-sensor camera imaging |
CN107451545A (en) * | 2017-07-15 | 2017-12-08 | 西安电子科技大学 | The face identification method of Non-negative Matrix Factorization is differentiated based on multichannel under soft label |
CN107832711A (en) * | 2017-11-13 | 2018-03-23 | 常州大学 | A kind of recognition methods again of the pedestrian based on transfer learning |
CN108230322A (en) * | 2018-01-28 | 2018-06-29 | 浙江大学 | A kind of eyeground feature detection device based on weak sample labeling |
CN109344787A (en) * | 2018-10-15 | 2019-02-15 | 浙江工业大学 | A kind of specific objective tracking identified again based on recognition of face and pedestrian |
Non-Patent Citations (3)
Title |
---|
WEITAO WAN ET AL: "Rethinking Feature Distribution for Loss Functions in Image Classification", 《CVPR2018》 * |
XINWEI HE ET AL: "Triplet-Center Loss for Multi-View 3D Object Retrieval", 《2018 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
乔治: "道路监控场景下行人再识别方法研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110968735A (en) * | 2019-11-25 | 2020-04-07 | 中国矿业大学 | Unsupervised pedestrian re-identification method based on spherical similarity hierarchical clustering |
CN111091548A (en) * | 2019-12-12 | 2020-05-01 | 哈尔滨市科佳通用机电股份有限公司 | Railway wagon adapter dislocation fault image identification method and system based on deep learning |
CN111666823A (en) * | 2020-05-14 | 2020-09-15 | 武汉大学 | Pedestrian re-identification method based on individual walking motion space-time law collaborative identification |
CN112818755A (en) * | 2021-01-13 | 2021-05-18 | 华中科技大学 | Gait recognition method based on active learning |
CN112818755B (en) * | 2021-01-13 | 2022-05-06 | 华中科技大学 | Gait recognition method based on active learning |
Also Published As
Publication number | Publication date |
---|---|
CN110309810B (en) | 2021-08-17 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111126360B (en) | Cross-domain pedestrian re-identification method based on unsupervised combined multi-loss model | |
CN106096561B (en) | Infrared pedestrian detection method based on image block deep learning features | |
Zhao et al. | Learning mid-level filters for person re-identification | |
CN110309810A (en) | A kind of pedestrian's recognition methods again based on batch center similarity | |
CN103207898B (en) | A kind of similar face method for quickly retrieving based on local sensitivity Hash | |
AU2010322173B2 (en) | Automatically mining person models of celebrities for visual search applications | |
WO2020206850A1 (en) | Image annotation method and device employing high-dimensional image | |
CN111310662B (en) | Flame detection and identification method and system based on integrated deep network | |
CN102880729A (en) | Figure image retrieval method and device based on human face detection and recognition | |
CN106462728A (en) | Systems and methods for image-feature-based recognition | |
CN102209975A (en) | Method for acquiring region-of-interest and/or cognitive information from eye image | |
CN114973317B (en) | Pedestrian re-recognition method based on multi-scale adjacent interaction characteristics | |
CN106295532B (en) | A kind of human motion recognition method in video image | |
CN105809205B (en) | A kind of classification method and its system of high spectrum image | |
CN107292642A (en) | A kind of Method of Commodity Recommendation and system based on image | |
CN102385592A (en) | Image concept detection method and device | |
CN112465854A (en) | Unmanned aerial vehicle tracking method based on anchor-free detection algorithm | |
CN115062186B (en) | Video content retrieval method, device, equipment and storage medium | |
CN110750663B (en) | Cross-modal image retrieval method for life records | |
CN106250396B (en) | Automatic image label generation system and method | |
Sun et al. | Indiscernible object counting in underwater scenes | |
CN110348366B (en) | Automatic optimal face searching method and device | |
CN109241315B (en) | Rapid face retrieval method based on deep learning | |
CN113628252A (en) | Method for detecting gas cloud cluster leakage based on thermal imaging video | |
Li et al. | Deep fusion of multi-layers salient CNN features and similarity network for robust visual place recognition |
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 |