CN105205501B - A kind of weak mark image object detection method of multi classifier combination - Google Patents
A kind of weak mark image object detection method of multi classifier combination Download PDFInfo
- Publication number
- CN105205501B CN105205501B CN201510643148.1A CN201510643148A CN105205501B CN 105205501 B CN105205501 B CN 105205501B CN 201510643148 A CN201510643148 A CN 201510643148A CN 105205501 B CN105205501 B CN 105205501B
- Authority
- CN
- China
- Prior art keywords
- grader
- classification
- class
- image object
- objectivity
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention is a kind of weak mark image object detection method of multi classifier combination, including：The weak mark image set of the different labels of input M carries out objectivity analysis to all pictures therein, generates objectivity region collection；To region collection generate characteristics of image, after different label characteristics collection are clustered respectively；According to cluster as a result, training media areas grader to each cluster rear region set；Each grader calculates separately category attribute；Input test image carries out objectivity and analyzes to obtain region unit, formation zone feature.Joint-detection is carried out using multi-categorizer, judges the region for including category object.The present invention has good performance in terms of multi-class image object joint-detection, can be applied to image object automatic marking, the fields such as image object identification.
Description
Technical field
The invention belongs to image procossing, technical field of computer vision, are a kind of weak mark images of multi classifier combination
Method for checking object.
Background technology
1, the object detection technique based on weak mark is mainly considered how using simple markup information and is not marked largely
The problem of sample is trained and classifies can preferably utilize a large amount of data to obtain relatively preferable on a low-cost basis
Recognition effect.In 2010, Alexe et al. proposed the concept of image object, was existed using the methods of notable Analysis on Prospect
There is no extraction on the image of any mark that may include the region of object, this method has considered color contrast (Color
Contrast), marginal density (Edge Density) and super-pixel span (Superpixels Straddling).
A kind of Weakly supervised study based on attribute knowledge and localization method, this method are proposed in Thomas in 2012 et al.
Some pictures for passing through weak mark are provided first, and object location information does not provide, and Weakly supervised study needs to learn an object
Body class models can be used to determine whether a test pictures even are located out comprising a classification
(binding-box).Location model in method is a dense CRF model.Wherein each training picture is a section
Point, node space size are enough comprising the window collection in picture.What single-point potential energy (unary potential) therein measured is
One window includes the possibility of a kind of type objects, and is then indicated for putting potential energy (pairwise potential) in pairs
Whether two windows include the object of same unknown classification.Method simultaneously carries out object by a kind of Special display model of study
Positioning, and is tested on different data sets, it was demonstrated that for many classifications has versatility.Based on weak mark
Object detection technique has quite extensive purposes in terms of large-scale image data processing, but in dividing to weak mark sample
When analysis, sample noise information can the final detection result of strong influence.
2, cluster is assumed to be the basic skills used in Weakly supervised study, and cluster hypothesis refers to the characteristic distance of sample data
When closer, sample may more belong to same category, and the example being in identical cluster (cluster) has larger possibility to possess
Identical label.1967, James MacQuee proposed k-means algorithms, give one group of observation data (x1, x2 ...,
Xn), wherein each observation data indicate that the purpose of k-means algorithms is that n observation is divided into k class with the vector that d is tieed up
Not：S={ S1,S2,…,Sk}.The formula of k-means algorithms is as follows：
According to cluster it is assumed that decision boundary just should be as possible by the more sparse place of data, to avoid dense
Cluster in data point assign to decision boundary both sides.Under this hypothesis, a large amount of unmarked exemplary effects are just to aid in spy
The dense and sparse region of data distribution in bright instance space, to which guidance learning algorithm has label to learn from example utilization
Decision boundary is adjusted, and makes it as possible by the sparse region of data distribution.
3, TF-IDF is a kind of statistical method, to assess words in a file set or a corpus wherein
The significance level of text document.Nineteen eighty-three, Salton et al. is in the works about text retrieval technique, it is proposed that TF-IDF's
Criterion shows the directly proportional increase of number that the importance of words occurs hereof with it, but simultaneously can be as it is in language material
The frequency occurred in library is inversely proportional decline.The main thought of TF-IDF is：If some word or phrase occur in an article
Frequency TF high, and seldom occur in other articles, then it is assumed that this word or phrase have good class discrimination ability,
It is adapted to classify.Word frequency (term frequency, TF) refers to the frequency that some given word occurs in this document
Rate.Reverse document-frequency (inverse document frequency, IDF) is the measurement of a word general importance.
In the Weakly supervised detection method of multi classifier combination, it will be assumed that pairwise classification utensil has similarity measurement, according to TF-IDF's
Concept, we can calculate the class of any grader by defining the similarity measurement in a certain grader and class between grader
Other singularity and classification independence.
Invention content
The purpose of the present invention is to propose to a kind of weak mark image object detection method of multi classifier combination, this methods
There is good performance on standard testing data set.
To complete the purpose of the present invention, the technical solution adopted by the present invention is：
A kind of weak mark image object detection method of multi classifier combination, wherein include the following steps：
Step (1), pictures pretreatment：The image data set of weak mark of the input comprising M class label, carries out object
Property is analyzed to obtain objectivity region collection (box proposals)：Weak mark pictures are given, including M classification, each classification mark
Label are defined as L={ L1,L2,...,Li(i=1,2 ..., M), objectivity analysis is carried out to all pictures therein, generates figure
The picture region set of blocks of hundred times of orders of magnitude of piece number.For each weak mark training set, the only mark of image collection classification
Label, and lack the markup information to image internal object position.
Step (2), formation zone feature carry out feature clustering according to different classes of label：To each region unit, we make
The convolutional neural networks model trained with ImageNet data sets extracts the feature (4096 dimension) of the fc7 layers of neural network model
As representing feature.
Step (3) each gathers each objectivity region collection training media areas grader according to cluster result
Cluster result to several cluster centres and to all areas feature, and all regions are divided by difference according to cluster result
Set.According to cluster result before, Linear SVM grader is respectively trained to above-mentioned each category set, is obtainedIt is a
Image media areas grader.Make whole features of this provincial characteristics collection of table as positive sample when the training of each middle level grader
This, and the feature selected immediately using in other characteristic sets is as negative sample, in specific training process, we utilize 10 times
10 folding cross validations (10-fold cross validation) establish more reliable and stable model.
Step (4) calculates each media areas grader category attribute：Each grader calculates separately and other graders
Correlation, classification singularity (category-specific attribute) and the classification for thus obtaining grader be unrelated
Property (category-irrelevant attribute).For the correlation of multi-categorizer, according to multi-categorizer analysis result,
It needs to calculate the similitude between grader and Conjoint Analysis and detection is carried out to input picture.In the similitude for calculating grader
When, we are firstly the need of the similitude between two cluster set of estimation：First life is closed in verification set validation collection
At objectivity detection zone.Then it is detected on validation set using each middle level grader, according to confidence level
Descending sequence is carried out to test result, sequencing numbers carry out being denoted as P (i, k), the highest region of Tr confidence level before taking, than
To its registration, the calculating formula of registration is as follows：
Wherein P (i, k) indicates that test result of i-th of grader in k-th of classification, M indicate the total number of classification.
It, respectively can be by i-th grader of being calculated of similar degree in the class and similar degree in the class after the similitude for calculating grader
Classification singularity T (i) (category-specific attribute) and classification independence D (i) (category-
irrelevant attribute).Category set where i-th of grader is denoted as gi, then have：
N(i,gi) indicating adjoining grader set of the i grader where itself in classification, T (i) is set N (i, gi)
The classification singularity of i-th interior of grader, Tc indicate grader number in set.N(i,k)(k≠gi) indicate j grader
In addition to classification where itself, the adjoining grader collection of remaining category set.D (i) is the classification independence of i-th of grader.T
(i) indicate that intercommunity of the grader in class, the grader for possessing larger T (i) have larger similar degree in the class, it can be by
Think preferably represent this class another characteristic.D (i) indicates intercommunity of the grader between different classes of, possesses
The grader of larger D (i) has similarity between larger class, is considered the identical field that can more show in different classes of object
Scenic spot characteristic of field, such as some common backgrounds.
Step (5) obtains combined detector according to multi classifier combination analysis result, to test pictures set (test
Set Conjoint Analysis and detection) are carried out.Objectivity analysis is equally carried out to the test image of input and obtains region unit, and is generated
Corresponding feature.Then we carry out corresponding joint test using multi-categorizer attributive analysis result, detect the meter of score
Formula：
In formula, v (i) is detection scores of the region R on i-th of grader, and v (i), D (i) are considered i-th
The word frequency attribute of grader and reverse document-frequency attribute.Final score can be considered as with the feature that has the category exclusive and
Eliminate common trait between class.Must itemize F (R) calculation：
F (R)=F0(R)+λ·o(R)
Wherein in order to improve the effect of joint-detection, we add objectivity estimation item o (R) in it must itemize, to more
Good improvement score estimation effect, wherein λ is 0 to 1 value, and λ, which is added, can effectively adjust the ratio of raw score item so that for
The recognition performance of object reaches best.Final result chooses ranking near preceding region as testing result.
Description of the drawings
Fig. 1 is the flow chart of the present invention；
Fig. 2 is the present invention for the positive and negative intuitive schematic diagram of grader cooperation detection；
Fig. 3 is the attributive analysis figure for multi-categorizer in the present invention；
Fig. 4 is the distribution map of multi-categorizer classification uniqueness and classification independence in the present invention；
Fig. 5 is the detection result figure of the present invention；
Fig. 6 is the joint detection process schematic diagram of the present invention.
Specific implementation mode
The present invention proposes a kind of weak mark image object detection method of multi classifier combination, includes the following steps：
(1) the multi-class trained pictures for including m classification are given and each give a class label to each classification,
It is defined as L={ L1,L2,...,Li(i=1,2 ..., M).Each classification includes NiImage (i=1,2 ..., M).
(2) objectivity analysis (Objectness Measure) is carried out to all images using objectivity detection method, often
It opens image and generates K candidate region.
(3) to each region unit, trained convolutional neural networks (CNN) mould on ImageNet data sets is utilized
Type, for each 4096 dimensional feature of region extraction fc7 layers.
(4) the target area feature generated to each classification is gathered using clustering algorithm (such as Kmeans algorithms) respectively
Class obtains C to the i-th classificationi=[Ni/ 100] a cluster centre and the cluster result to all areas feature.
(5) Linear SVM grader is respectively trained to above-mentioned each category set, using 10- in the cluster result before basis
Fold cross-validation methods be iterated optimization (data set is divided into 10 subsets, 10 sons concentrate 1 is used as test set,
And remaining 9 data set is as training set), it obtainsA image media areas grader.The instruction of each middle level grader
Using whole features of this provincial characteristics collection as positive sample when practicing, and use the feature selected immediately in other characteristic sets
As negative sample.
(6) similitude between grader is calculated, the similitude between two cluster set of estimation is needed：It is verifying first
Set validation collection, which closes, generates objectivity detection zone.Then utilize each middle level grader in validation set
On be detected, descending sequence is carried out to test result according to confidence level, sequencing numbers carry out being denoted as P (i, k), take preceding Tr
A highest region of confidence level, compares its registration, and the calculating formula of registration is as follows：
Wherein P (i, k) indicates that test result of i-th of grader in k-th of classification, M indicate the total number of classification.
(7) as shown in Fig. 2, need calculate grader similitude after, respectively can by class between class similarity measurements
Measure the classification unique (TF) and classification independence (DF) of grader.The grader set of i classes is denoted as gi, then have：
Rough T (i), D (i) distribution situations such as Fig. 4 are shown.
(8) the picture set of input is analyzed using objectivity and K candidate region is generated to every image, and with identical
Mode formation zone feature carries out joint-detection using multi-categorizer for each candidate region, detects the calculating formula of score：
(9) final score can be considered as eliminating common trait between class with the feature for having the category exclusive, that is, gather around
Have anisotropic between the same sex and class in minimum class.And in order to improve joint-detection effect we add objectivity in it must itemize and estimate
Meter：
F (R)=F0(R)+λ·o(R)
Final result chooses ranking near preceding region as testing result.
Claims (8)
1. a kind of weak mark image object detection method of multi classifier combination, it is characterised in that include the following steps：
(1) pictures pre-process：The image data set of weak mark of the input comprising M class label carries out objectivity and analyzes
To objectivity region collection (box proposals)；
(2) formation zone feature carries out feature clustering according to different classes of label；
(3) according to cluster result, to each objectivity region collection training media areas grader；
(4) each media areas grader category attribute is calculated；
(5) according to multi classifier combination analysis result, combined detector is obtained, is detected for image object.
2. the weak mark image object detection method of multi classifier combination as described in claim 1, it is characterised in that：Wherein,
In the step (1), the input training pictures of selection include M classification, and each classification includes NiImage, i=1,
2,...,M；Objectivity analysis is carried out to all images using objectivity detection method, every image generates K candidate region；M
The class label collection of a classification is defined as L={ L1,L2,...,Li(i=1,2 ..., M), for each weak mark training set,
The class label for only needing input picture set, without inputting the specific markup information to image internal object position.
3. the weak mark image object detection method of multi classifier combination as described in claim 1, which is characterized in that wherein institute
Step (2) is stated, when carrying out signature analysis for all image candidate regions of generation, the volume trained using ImageNet data sets
Neural network model is accumulated, the feature conduct for extracting the fc7 layers of neural network model represents feature totally 4096 dimension；Later to each class
The target area feature not generated is clustered using clustering algorithm respectively, and C is obtained to the i-th classificationi=[Ni/ 100] in a cluster
The heart and cluster result to all target area features.
4. the weak mark image object detection method of multi classifier combination as claimed in claim 3, it is characterised in that：The step
Suddenly (3) are respectively trained Linear SVM grader to above-mentioned each category set, obtain according to cluster result beforeA figure
As media areas grader；Using whole features of this provincial characteristics collection as positive sample when the training of each middle level grader
This, and the feature selected at random using in other characteristic sets is as negative sample.
5. the weak mark image object detection method of multi classifier combination as described in claim 1, it is characterised in that：The step
Suddenly (4) calculate the correlation with other graders in the same category, obtain grader for each media areas grader
Classification singularity；For each media areas grader, the correlation of each grader and other different classes of graders is calculated,
Obtain the classification independence of grader.
6. the weak mark image object detection method of multi classifier combination as claimed in claim 5, it is characterised in that：For more
The correlation of grader needs to calculate the similitude between grader and is carried out to input picture according to multi-categorizer analysis result
Conjoint Analysis and detection；When calculating the similitude of grader, it is necessary first to estimate the similitude between two cluster set：First
It is closed in verification collection and generates objectivity detection zone；Then it is detected in verification set using each middle level grader, according to
Descending sequence is carried out to test result according to confidence level, sequencing numbers carry out being denoted as P (i, k), Tr confidence level highest before taking
Region, compare its registration, the calculating formula of registration is as follows：
Wherein P (i, k) indicates that test result of i-th of grader in k-th of classification, M indicate the total number of classification.
7. the weak mark image object detection method of multi classifier combination as claimed in claim 5, it is characterised in that：It is calculating
After the similitude of grader, the classification singularity (category- of grader is obtained by similitude in similitude between class and class respectively
Specific attribute) and classification independence (category-irrelevant attribute)；By the classification of i classes
Device set is denoted as gi, then have：
T (i) is classification giThe classification singularity of i-th interior of grader, D (i) are classification giThe classification of i-th interior of grader
Independence；T (i) indicates similarity of the grader in class, possesses the grader of larger T (i), is considered to better generation
This class another characteristic of table；T (i) indicates similarity of the grader between class, possess larger D (i) grader have it is larger
Class between similarity, it is considered to be can more show the same scene provincial characteristics in different classes of object.
8. the weak mark image object detection method of multi classifier combination as described in claim 1, it is characterised in that：It obtains
Multiple media areas graders and corresponding attributive analysis are as a result, rear carry out joint test using multi-categorizer；For defeated
The test pictures set entered similarly generates K candidate region, and formation zone feature in the same manner, for each candidate region
Joint-detection is carried out using multi-categorizer, detects the calculating formula of score：
In formula, v (i) is detection scores of the region R on i-th of grader, and v (i), D (i) are considered i-th of classification
The word frequency attribute of device and reverse document-frequency attribute；Final score is considered as possessing the exclusive feature of the category and eliminating
Common trait between class possesses in minimum class diversity between similitude and class；Must itemize F (R) calculation：
F (R)=F0(R)+λ·o(R)
And in order to improve joint-detection effect we added in it must itemize objectivity estimation, wherein λ is 0 to 1 value, be added
λ can effectively adjust final result so that reach best for the recognition performance of object；Final result chooses ranking near preceding
Region is as testing result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510643148.1A CN105205501B (en) | 2015-10-04 | 2015-10-04 | A kind of weak mark image object detection method of multi classifier combination |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510643148.1A CN105205501B (en) | 2015-10-04 | 2015-10-04 | A kind of weak mark image object detection method of multi classifier combination |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105205501A CN105205501A (en) | 2015-12-30 |
| CN105205501B true CN105205501B (en) | 2018-09-18 |
Family
ID=54953173
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510643148.1A Active CN105205501B (en) | 2015-10-04 | 2015-10-04 | A kind of weak mark image object detection method of multi classifier combination |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105205501B (en) |
Families Citing this family (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108496185B (en) * | 2016-01-18 | 2022-09-16 | 北京市商汤科技开发有限公司 | System and method for object detection |
| CN108268510B (en) * | 2016-12-30 | 2022-01-28 | 华为技术有限公司 | Image annotation method and device |
| CN106886995B (en) | 2017-01-13 | 2019-09-20 | 北京航空航天大学 | Polyteny example returns the significant object segmentation methods of image of device polymerization |
| US11216927B2 (en) * | 2017-03-16 | 2022-01-04 | Siemens Aktiengesellschaft | Visual localization in images using weakly supervised neural network |
| CN108009279B (en) * | 2017-12-20 | 2023-02-17 | 华东理工大学 | Image area label correction method based on spatial position relation graph matching |
| CN108171191B (en) * | 2018-01-05 | 2019-06-28 | 百度在线网络技术（北京）有限公司 | Method and apparatus for detecting face |
| CN108416384B (en) * | 2018-03-05 | 2021-11-05 | 苏州大学 | Image label labeling method, system, equipment and readable storage medium |
| CN108734195A (en) * | 2018-04-13 | 2018-11-02 | 王延峰 | Weakly supervised detection model training method based on Cooperative Study and system |
| CN109359683B (en) * | 2018-10-15 | 2021-07-27 | 百度在线网络技术（北京）有限公司 | Target detection method, device, terminal and computer-readable storage medium |
| CN109325631A (en) * | 2018-10-15 | 2019-02-12 | 华中科技大学 | Electric car charging load forecasting method and system based on data mining |
| CN109543693B (en) * | 2018-11-28 | 2021-05-07 | 中国人民解放军国防科技大学 | Weak labeling data noise reduction method based on regularization label propagation |
| CN111061890B (en) * | 2019-12-09 | 2023-04-07 | 腾讯云计算（北京）有限责任公司 | Method for verifying labeling information, method and device for determining category |
| CN111738354A (en) * | 2020-07-20 | 2020-10-02 | 深圳市天和荣科技有限公司 | Automatic recognition training method, system, storage medium and computer equipment |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2007513407A (en) * | 2003-11-12 | 2007-05-24 | コーニンクレッカ フィリップス エレクトロニクス エヌ ヴィ | Adding semantic indicators to phrases for grammar generation |
| EP2434434A2 (en) * | 2010-09-22 | 2012-03-28 | Siemens Corporation | Method and system for training a landmark detector using multiple instance learning |
| CN103310195A (en) * | 2013-06-09 | 2013-09-18 | 西北工业大学 | LLC-feature-based weak-supervision recognition method for vehicle high-resolution remote sensing images |
| CN104573669A (en) * | 2015-01-27 | 2015-04-29 | 中国科学院自动化研究所 | Image object detection method |
-
2015
- 2015-10-04 CN CN201510643148.1A patent/CN105205501B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2007513407A (en) * | 2003-11-12 | 2007-05-24 | コーニンクレッカ フィリップス エレクトロニクス エヌ ヴィ | Adding semantic indicators to phrases for grammar generation |
| EP2434434A2 (en) * | 2010-09-22 | 2012-03-28 | Siemens Corporation | Method and system for training a landmark detector using multiple instance learning |
| CN103310195A (en) * | 2013-06-09 | 2013-09-18 | 西北工业大学 | LLC-feature-based weak-supervision recognition method for vehicle high-resolution remote sensing images |
| CN104573669A (en) * | 2015-01-27 | 2015-04-29 | 中国科学院自动化研究所 | Image object detection method |
Non-Patent Citations (5)
| Title |
|---|
| Automatic image annotation and retrieval using cross-media relevance models;J. Jeon等;《SIGIR "03: Proceedings of the 26th annual international ACM SIGIR conference on Research and development in informaion retrieval》;20030801;119-126 * |
| Automatic image annotation via local multi-label classification;Mei Wang等;《CIVR "08 Proceedings of the 2008 international conference on Content-based image and video retrieval》;20080709;17-26 * |
| Automatic image annotation with weakly labeled dataset;Wei Zhang等;《MM "11 Proceedings of the 19th ACM international conference on Multimedia》;20111201;1185-1188 * |
| Learning a Part-Based Pedestrian Detector in a Virtual World;Xu Jiaolong等;《IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS》;20141031;第15卷(第5期);2121-2131 * |
| Multi-level annotation of natural scenes using dominant image components and semantic concepts;Jianping Fan等;《MULTIMEDIA "04 Proceedings of the 12th annual ACM international conference on Multimedia》;20041016;540-547 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105205501A (en) | 2015-12-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105205501B (en) | A kind of weak mark image object detection method of multi classifier combination | |
| Hendricks et al. | Women also snowboard: Overcoming bias in captioning models | |
| CN109086658B (en) | Sensor data generation method and system based on generation countermeasure network | |
| Socarrás et al. | Adapting pedestrian detection from synthetic to far infrared images | |
| CN109101938B (en) | Multi-label age estimation method based on convolutional neural network | |
| CN105787472B (en) | A kind of anomaly detection method based on the study of space-time laplacian eigenmaps | |
| CN106228142A (en) | Face verification method based on convolutional neural networks and Bayesian decision | |
| CN106407958B (en) | Face feature detection method based on double-layer cascade | |
| CN102436589B (en) | Complex object automatic recognition method based on multi-category primitive self-learning | |
| Bansal et al. | Towards transparent systems: Semantic characterization of failure modes | |
| CN105389326B (en) | Image labeling method based on weak matching probability typical relevancy models | |
| Li et al. | HEGM: A hierarchical elastic graph matching for hand gesture recognition | |
| CN109002834A (en) | Fine granularity image classification method based on multi-modal characterization | |
| CN103824051A (en) | Local region matching-based face search method | |
| CN108960142B (en) | Pedestrian re-identification method based on global feature loss function | |
| CN110503000B (en) | Teaching head-up rate measuring method based on face recognition technology | |
| CN101799875B (en) | Target detection method | |
| CN108492298A (en) | Based on the multispectral image change detecting method for generating confrontation network | |
| CN106203483A (en) | A kind of zero sample image sorting technique of multi-modal mapping method of being correlated with based on semanteme | |
| CN107194413A (en) | A kind of differentiation type based on multi-feature fusion cascades the target matching method of display model | |
| Iqbal et al. | Mitochondrial organelle movement classification (fission and fusion) via convolutional neural network approach | |
| CN105205807B (en) | Method for detecting change of remote sensing image based on sparse automatic coding machine | |
| CN105117703A (en) | Rapid action unit recognition method based on matrix multiplication | |
| CN106599834A (en) | Information pushing method and system | |
| Chen et al. | Patch-based topic model for group detection |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |