CN104573744B - Fine granulation classification identifies and the part of object positions and feature extracting method - Google Patents
Fine granulation classification identifies and the part of object positions and feature extracting method Download PDFInfo
- Publication number
- CN104573744B CN104573744B CN201510026025.3A CN201510026025A CN104573744B CN 104573744 B CN104573744 B CN 104573744B CN 201510026025 A CN201510026025 A CN 201510026025A CN 104573744 B CN104573744 B CN 104573744B
- Authority
- CN
- China
- Prior art keywords
- detector
- feature
- partial
- detection
- characteristic pattern
- 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
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (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)
- Probability & Statistics with Applications (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention proposes a kind of identification of fine granulation classification and the part positioning of object and feature extracting method, this method preferably solve the problems, such as the part orientation problem and feature representation problem of object in fine granulation classification identification.For the part orientation problem of object, a series of partial detector is trained using supervised learning, in view of positioning the attitudes vibration and influence of crust deformation of target, this method only detects the smaller part of deformation, and the same object parts are trained with different detectors using posture clustering method, to which the attitudes vibration of object is taken into account.For object or partial feature representation, this method proposes to extract feature in multiple scales and multiple positions, then merges these features for final object representation, so that this feature has certain scale and translation invariance.There is certain complementarity simultaneously between the object parts positioning of the present invention and feature representation, so as to effectively improve the precision that fine classification identifies problem.
Description
Technical field
The present invention relates to a kind of methods of technical field of image processing, specifically, referring to a kind of fine granulation class
The part of object involved in other recognition methods and the identification problem positions and feature extracting method.
Background technology
The target of fine granulation classification problem is to discriminate between hundreds of multiple subclass under same major class, such as area
Divide different classes of flower, bird, dog etc..For layman, identify that these subclasses are very difficult, fine granulations
The it is proposed of classification problem solves the problems, such as that layman identifies these similar subclasses.User only needs given target object,
Pass through fine classification recognition methods, so that it may to return to the classification of target object, and then the subclass series of characteristics can be obtained.No
It is same as general category identification problem (such as distinguishing car and people), since the comparison in difference between subclass is small and height part
Change, it is very difficult to distinguish these subclasses.It is widely used in the spatial pyramid model of general category identification problem due to not
This high localized sub- class inherited can be captured, thus satisfied recognition result cannot be reached.
By the literature search discovery to the prior art, the difficult point of fine granulation classification problem is mainly at two aspects, i.e.,
Part positioning and iamge description.It was P.Felzenszwalb in 2010 that part, which positions widely used,《IEEE
Transactions on Pattern Analysis and Machine Intelligence》On the ``A that delivers
Discriminatively trained, multiscale, deformable part model ", you can deformed part model with
And its mutation.The model finds target object or partial target object by training template detector, and considers
Geometric relativity between department pattern.However, only the part smaller to deformation has preferable detection result to the model,
The part bigger to deformation, the wing of such as bird, the poor performance of part detection model.For iamge description, mostly use greatly
D.G.Lowe was published in 2004《International Journal of Computer Vision》On ``
Distinctive image features from scale-invariant keypoints ", i.e. scale invariant feature.So
And this feature is only the combination of some gradient informations, independently of specific data set, does not have preferable separating capacity.Its
His feature such as Krizhevsky was published in 2010《Neural Information Processing Systems》On
``Imagenet classification with deep convolutional neural networks ", i.e., volume and god
Through network characterization, this feature lacks enough scales and puts down despite the abundant feature of the semanteme for design data
Motion immovability.When the fractional object of detection and actual position have relatively large deviation, this feature cannot overcome this translation well
Variation.
Invention content
For the defects in the prior art, the object of the present invention is to provide a kind of identification of fine granulation classification and its objects
Part positioning and feature extracting method improve the precision of part positioning and the scale invariability and translation invariant of feature representation
Property, to improve the accuracy of identification of fine classification classification problem.
The present invention is achieved by the following technical solutions:
According to the first aspect of the invention, a kind of part localization method of object is provided, is that one kind is divided for fine granulation
Sector of breakdown localization method, this method are smaller using object detector and partial detector detection target object and its deformation
Part, the detector are that have measure of supervision to learn using what posture clustered, it is contemplated that object or partial appearance
State changes;Object detector and partial detector independently carry out, and return to the detection zone that score is high in each detector
As candidate, final testing result is obtained by correcting object and partial detection.
Preferably, the detector is that have measure of supervision to learn using what posture clustered, specially:For object
And each part, positive example sample is assembled to some mixed models according to posture;
Assuming that each part piAll with a bounding boxDefinition, whole object is with bounding box p0
Expression, wherein (l, t, r, b) indicates the left side of bounding box, top, right side and bottom coordinate position;By following vectorial, this
The part demarcated a bit is used for parameterizing the posture θ of sample II:
θI=(p '1, p '2..., p 'n)
Wherein, w and h indicates object p0Width and height, n indicate object parts quantity .p 'iIt is piNormalization table
It reaches, this normalized expression allows to only consider the relative position of part, and ignores the ruler between different objects part
Spend difference;All positive samples are clustered into C ingredient according to attitude characteristic using k- means clustering methods.
Further, it is possible to inconsistent with the position of object to solve object parts in the testing result returned, it is described
Object detector and partial detector return to the detection zone that score is high in each detector and are used as candidate, specially:
Enable X={ x0, x1..., xnIndicate the high testing result of the score of object and its corresponding n part, φ (X)=
{φ(x0), φ (x1) ..., φ (xn) indicate corresponding convolution feature, a series of detector { w that given training obtains0,
w1..., wn, update testing result by optimizing following expression:
Wherein
Ψ [] is a nonlinear function, and detection score is mapped to range [- 1,1], []εIt is a loss letter
Number;Parameter lambdaiThe degree of overlapping of measurement part and object, ranging from [0,1];Weighted term [λi]εFor punishing part and the object of detection
The inconsistent situation of body.
Part positioning of the present invention is only used for the smaller part of object deformation, and the training of detector considers aspect
Variation, the object by geometry update detection and its partial relationship, obtain reliable object and its partial positioning accuracy.
According to the second aspect of the invention, a kind of feature extracting method is provided, this method is positioned in each object parts and tied
Constant convolution feature is extracted on fruit, i.e., extracts convolution feature in multiple scales and multiple visual angles, these convolution features carry out
Fusion obtains final feature representation, which is used for final classification.
Further, the feature extracting method, includes the following steps:
Step 1:To given scale image, the 5th convolutional layer characteristic pattern f is extractedw×h×C, wherein w × h represents trellis diagram
As size, C represents the port number of characteristic pattern;Input picture is 16 relative to the down-sampling ratio of the 5th convolutional layer, it is meant that
5th convolutional layer characteristic pattern is 16 relative to the step-length of input picture;
Step 2:Zero padding operation is carried out to the boundary of each channel of characteristic pattern, increases by two pixels per side, thus
The characteristic pattern f ' later to zero paddingw′×h′×C;The characteristic pattern f ' later to zero paddingw′×h′×C, sliding window is used on each channel
Method selects arbitrary subgraph f with step-length 1w×h×C, therefore a total of 5 × 5 relative to the upper left corner biasing (Δ x, Δ y) be 0,1,2,
3,4 } subgraph;Then pondization operation is carried out to each subgraph, obtains target output size as the later son of the pondization of n × n
Figure;
Step 3:Use the full unicom layer characteristic pattern of pond beggar's figure calculated for subsequent obtained in step 2.
Preferably, aforesaid operations carry out on 5 scales of input picture and its flip horizontal image, are finally always obtained
25 × 5 × 2 feature vectors, these feature vectors carry out obtaining on each scale after pondization operation on each scale respectively
Single features expression, finally cascade feature on these multiple scales for the final expression to image.This feature makes it
With certain scale and translation invariance.
According to the third aspect of the invention we, a kind of raising image fine granulation knowledge method for distinguishing is provided, is included the following steps:
The first step:For test image, object detector and partial detector detection target object and its deformation are utilized
Small part, the detector are that have measure of supervision to learn using what posture clustered, it is contemplated that object is partial
Attitudes vibration.Since object detector and partial detector independently carry out, the geometrical relationship between them is not accounted for.
Lead to as candidate, final testing result as an improvement, this method returns to some high detection zones of score in each detector
Overcorrect object and partial detection obtain.
Second step rolls up the object or part that are each detected in the first step in multiple scales and the extraction of multiple visual angles
Product feature, these convolution features are merged to obtain final feature representation, which is used for final classification.The present invention can carry
Hi-vision fine granulation identifies.
To sum up, the method for the present invention preferably solve the problems, such as in fine granulation classification identification the part orientation problem of object and
Feature representation problem, the method increase part detection performances, and this feature is made to have certain scale and translation invariant
Property.There is certain complementarity simultaneously between the object parts positioning of the present invention and feature representation, so as to effectively improve
Fine classification identifies the precision of problem.
Compared with prior art, the present invention has following advantageous effect:
Above-mentioned technical proposal of the present invention preferably solves the problems, such as that the part positioning of object in fine granulation classification identification is asked
Topic and feature representation problem.The currently volume with preferable ability to express is all utilized in the part positioning of the present invention and feature representation
Product neural network.Present invention employs the strong supervised learning method training objective detectors clustered based on posture, and to final
Testing result carries out geometry update, can obtain accurate part positioning accuracy.Meanwhile constant feature representation technology energy
Enough inaccuracies for overcoming positioning to a certain extent, make it have certain scale and translation invariance.Two methods
In conjunction with allowing the invention to obtain preferable recognition performance in fine granulation classification problem.
Description of the drawings
Upon reading the detailed description of non-limiting embodiments with reference to the following drawings, other feature of the invention,
Objects and advantages will become more apparent upon:
Fig. 1 is the principle framework figure of one embodiment of the invention;
Fig. 2 is the invariant feature extraction flow chart of one embodiment of the invention.
Specific implementation mode
With reference to specific embodiment, the present invention is described in detail.Following embodiment will be helpful to the technology of this field
Personnel further understand the present invention, but the invention is not limited in any way.It should be pointed out that the ordinary skill of this field
For personnel, without departing from the inventive concept of the premise, various modifications and improvements can be made.These belong to the present invention
Protection domain.
As shown in Figure 1, for the principle framework of one embodiment of the invention, which includes two parts, the i.e. part of object
Position portion and scale and translation invariant feature representation part.A given width test image, first with object detector
And partial detector detection target object and its deformation smaller portions, the detector are that have supervision using what posture clustered
Methodology acquistion is arrived, it is contemplated that object or partial attitudes vibration.Since object detector and partial detector are only
Vertical progress, do not account for the geometrical relationship between them.As an improvement, returning to the higher detection of score in each detector
Region is obtained as candidate, final testing result by correcting object and partial detection.Then, characteristic extracting module pair
The object or part each detected extracts convolution feature in multiple scales and multiple visual angles, these convolution features are merged
Final feature representation is obtained, which is used for final classification.
As a preferred embodiment, the object and part detection specific implementation process include the following steps:
Step 1:
Posture clusters:This method is using the learning method training detector supervised by force, for training sample, whole object with
And the bounding box of some object parts is all known.For object and each part, this method is assembled just according to posture
Example sample is to some mixed models.Assuming that each part piAll with a bounding boxDefinition is (entire
Object is with bounding box p0Expression), wherein (l, t, r, b) indicates the left side of bounding box, top, right side and bottom coordinate position.
By following vector, the part of these calibration is used for parameterizing the posture θ of sample II:
θI=(p '1, p '2..., p 'n)
Wherein, w and h indicates object p0Width and height, n indicate object parts quantity .p 'iIt is piNormalization table
It reaches, this normalized expression allows to only consider the relative position of part, and ignores the ruler between different objects part
Spend difference.All positive samples are clustered into C ingredient according to attitude characteristic using k- means clustering methods.The cluster is considered
The attitudes vibration of object, this is very important the training of detector.
Step 2:
Convolutional network training and detection son study:It is characterized in extracting from convolutional network for detector training.
In order to enable convolutional network adapts to specific fine classification data set, convolutional neural networks should be finely tuned first.Due to training sample
This is limited, generates a series of sub-district area images using selection searching method first, wherein all Chong Die with original positive example sample big
It is considered as positive example in 0.5 subsample, and every other subsample is considered as negative example, thus obtains finely tuning later convolution
Neural network.In detector training process, only only original sample feature is considered as positive example, those are Chong Die with original sample
Subsample of the degree less than 0.3 is considered as negative example.To each part of object and object, can be obtained with stand-alone training a series of
Detector { w0, w1..., wn}。
When test, a series of candidate sub-district equally is generated using the method for selection search to a width test image
Domain.The feature of each candidate subregion x is indicated with φ (x), is directed to detector w accordinglyiScore be then expressed as
Wherein the higher region of score (such as 100, which can be set as needed) are selected as couple candidate detection result.
Step 3:
Object and part detection update:Since the detection to object and object parts independently carries out, the inspection of return
It surveys object parts in result and is possible to inconsistent with the position of object.A kind of geometry update method be used to solve the problems, such as this.
Enable X={ x0, x1..., xnIndicate that (region quantity can for higher 100 regions of score of object and corresponding n part
To be set as needed) testing result, φ (X)={ φ(x0), φ (x1) ..., φ (xn) indicate corresponding convolution feature.
A series of detector { w that given training obtains0, w1..., wn, update testing result by optimizing following expression:
Wherein
Ψ [z] is a nonlinear function, and detection score is mapped to range [- 1,1], [λi]εIt is a loss function.
Parameter lambdaiThe degree of overlapping of measurement part and object, ranging from [0,1].Weighted term [λi]ε(ε=0.6) is used for punishing the part of detection
The inconsistent situation with object.
Characteristic extraction part is as shown in Fig. 2, test image is adjusted ratio to different scales, for each scale,
Extraction feature mainly includes the following steps:
Step 1:To given scale image, the 5th convolutional layer characteristic pattern f is extractedw×h×C, wherein w × h represents trellis diagram
As size, C represents the port number of characteristic pattern.Input picture is 16 relative to the down-sampling ratio of the 5th convolutional layer, it is meant that
5th convolutional layer characteristic pattern is 16 relative to the step-length of input picture.
Step 2:Zero padding operation is carried out to the boundary of each channel of characteristic pattern, increases by two pixels per side, thus
The characteristic pattern f ' later to zero paddingw′×h′×C.The characteristic pattern f ' later to zero paddingw′×h′×C, sliding window is used on each channel
Method selects arbitrary subgraph f with step-length 1w×h×C, therefore a total of 5 × 5 relative to the upper left corner biasing (Δ x, Δ y) be 0,1,2,
3,4 } subgraph.Then pondization operation is carried out to each subgraph, obtains target output size as the later son of the pondization of n × n
Figure.
Step 3:Use the full unicom layer characteristic pattern of pond beggar's figure calculated for subsequent obtained in step 2.
Aforesaid operations carry out on 5 scales of input picture and its flip horizontal image, therefore are finally always obtained 25
× 5 × 2 feature vectors.These feature vectors carry out obtaining on each scale after pondization operation on each scale respectively
Single features are expressed, and finally cascade the feature on these multiple scales for the final expression to image.
Implementation result:
Experiment carries out on widely used fine granulation data set CUB-200-2011.The data set include 200 not
Congener birds, in total 11788 width image.Identify that these subclasses are all very difficult for people.In deformation smaller part
In the selection divided, head and body are only chosen as part detection object.In the selection of posture cluster, each detection mesh
Mark is clustered into 3 mixed models, and when feature extraction, 5 scales are chosen for { 227,280,340,400,454 }.Final experiment
Standard is weighed with nicety of grading.
Object/part positioning accuracy result:
Positioning accuracy is weighed with the ratio being properly positioned, and the principle being properly positioned is that the target of detection is Chong Die with realistic objective
Degree is more than 0.5.To object, head and body part, this method can obtain 96.36%, 75.22%70.14%'s respectively
Position precision.
Classification results:
Have benefited from higher object/part positioning accuracy, the classifying identification method based on part can finally obtain
77.51% discrimination is far above existing accuracy of identification under same experiment condition.The validity of this method comes from
There are a complementary relationship in accurate part positioning accuracy and constant feature representation form, the two parts again simultaneously, i.e. feature
Invariance expresses the inaccuracy for compensating for positioning to a certain extent, further improves final image recognition precision.
Specific embodiments of the present invention are described above.It is to be appreciated that the invention is not limited in above-mentioned
Particular implementation, those skilled in the art can make various deformations or amendments within the scope of the claims, this not shadow
Ring the substantive content of the present invention.
Claims (4)
1. the part localization method of object in a kind of fine granulation classification identification, which is characterized in that this method utilizes object detection
Device and partial detector detection target object and its deformation smaller portions, the detector are that have prison using what posture clustered
Superintend and direct what methodology acquistion was arrived, it is contemplated that object or partial attitudes vibration;Object detector and partial detector are independent
It carries out, and returns to the detection zone that score is high in each detector and pass through correction object as candidate, final testing result
It is obtained with partial detection;
The detector is that have measure of supervision to learn using what posture clustered, specially:For object and each
Positive example sample is assembled to some mixed models in part according to posture;
Assuming that each part piAll with a bounding boxDefinition, whole object is with bounding box poExpression,
Wherein (l, t, r, b) indicates the left side of bounding box, top, right side and bottom coordinate position;Pass through following vector, these calibration
Part be used for parameterizing the posture θ of sample II:
θI=(p '1, p '2..., p 'n)
Wherein, w and h indicates object p0Width and height, n indicate object parts quantity, p 'iIt is piNormalization expression,
This normalized expression allows to only consider the relative position of part, and the scale ignored between different objects part is poor
It is different;All positive samples are clustered into C ingredient according to attitude characteristic using k- means clustering methods;
It is possible to inconsistent with the position of object to solve object parts in the testing result returned, the object detector and portion
Divide detector to return to the detection zone that score is high in each detector and be used as candidate, specially:
Enable X={ x0, x1..., xnIndicate the high testing result of the score of object and its corresponding n part, φ (X)={ φ
(x0), φ (x1) ..., φ (xn) indicate corresponding convolution feature, a series of detector { w that given training obtains0, w1...,
wn], update testing result by optimizing following expression:
Wherein
Ψ [] is a nonlinear function, and detection score is mapped to range [- 1,1], []∈It is a loss function;Ginseng
Number λiThe degree of overlapping of measurement part and object, ranging from [0,1];Weighted term [λi]∈For punishing that the part of detection is differed with object
The case where cause.
2. a kind of fine granulation classification recognition methods using 1 the method for the claims, it is characterised in that including two
Step:
The first step:For test image, using object detector and partial detector detection target object and its deformation compared with
Fraction, the detector are that have measure of supervision to learn using what posture clustered, it is contemplated that object is partial
Attitudes vibration;Object detector and partial detector independently carry out, and return to the detection zone that score is high in each detector
Domain is obtained as candidate, final testing result by correcting object and partial detection;
Second step, it is special in multiple scales and multiple visual angles extraction convolution to the object each detected in the first step or part
Sign, these convolution features are merged to obtain final feature representation, which is used for final classification.
3. fine granulation classification recognition methods according to claim 2, which is characterized in that the second step, including it is as follows
Step:
Step 1:To given scale image, the 5th convolutional layer characteristic pattern f is extractedw×h×C, it is big to represent convolved image wherein w × h
Small, C represents the port number of characteristic pattern;Input picture is 16 relative to the down-sampling ratio of the 5th convolutional layer, it is meant that the 5th
A convolutional layer characteristic pattern is 16 relative to the step-length of input picture;
Step 2:Zero padding operation is carried out to the boundary of each channel of characteristic pattern, increases by two pixels per side, is thus mended
Zero later characteristic pattern f 'w′×h′×C;The characteristic pattern f ' later to zero paddingw′×h′×C, on each channel using slip window sampling with
Step-length 1 selects arbitrary subgraph fw×h×C, therefore a total of 5 × 5 relative to the upper left corner biasing (Δ x, Δ y) be { 0,1,2,3,4 }
Subgraph;Then pondization operation is carried out to each subgraph, obtains target output size as the later subgraph of the pondization of n × n;
Step 3:Use the full unicom layer characteristic pattern of pond beggar's figure calculated for subsequent obtained in step 2.
4. fine granulation classification recognition methods according to claim 3, it is characterised in that:Above-mentioned steps operation is schemed in input
It is carried out on 5 scales of picture and its flip horizontal image, 25 × 5 × 2 feature vectors, these feature vectors is finally always obtained
It carries out obtaining the single features expression on each scale after pondization operation respectively on each scale, it is multiple finally to cascade these
Feature on scale is for the final expression to image.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510026025.3A CN104573744B (en) | 2015-01-19 | 2015-01-19 | Fine granulation classification identifies and the part of object positions and feature extracting method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510026025.3A CN104573744B (en) | 2015-01-19 | 2015-01-19 | Fine granulation classification identifies and the part of object positions and feature extracting method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN104573744A CN104573744A (en) | 2015-04-29 |
CN104573744B true CN104573744B (en) | 2018-07-20 |
Family
ID=53089763
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510026025.3A Active CN104573744B (en) | 2015-01-19 | 2015-01-19 | Fine granulation classification identifies and the part of object positions and feature extracting method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN104573744B (en) |
Families Citing this family (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107624189B (en) * | 2015-05-18 | 2020-11-20 | 北京市商汤科技开发有限公司 | Method and apparatus for generating a predictive model |
CN106097353B (en) * | 2016-06-15 | 2018-06-22 | 北京市商汤科技开发有限公司 | Method for segmenting objects and device, computing device based on the fusion of multi-level regional area |
CN106780479A (en) * | 2016-12-31 | 2017-05-31 | 天津大学 | A kind of high precision image fuzzy detection method based on deep learning |
CN107766890B (en) * | 2017-10-31 | 2021-09-14 | 天津大学 | Improved method for discriminant graph block learning in fine-grained identification |
CN108764247B (en) * | 2018-04-13 | 2020-11-10 | 中国科学院自动化研究所 | Dense connection-based deep learning object detection method and device |
CN108509939A (en) * | 2018-04-18 | 2018-09-07 | 北京大学深圳研究生院 | A kind of birds recognition methods based on deep learning |
CN109086792A (en) * | 2018-06-26 | 2018-12-25 | 上海理工大学 | Based on the fine granularity image classification method for detecting and identifying the network architecture |
CN111091150A (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 cross rod cover plate fracture detection method |
CN111177305B (en) * | 2019-12-30 | 2021-01-05 | 广州骏伯网络科技有限公司 | Method, system, device and storage medium for time segment selection |
CN112329603B (en) * | 2020-11-03 | 2022-09-13 | 西南科技大学 | Dam face crack defect positioning method based on image cascade |
CN112613371A (en) * | 2020-12-16 | 2021-04-06 | 上海大学 | Hyperspectral image road extraction method based on dense connection convolution neural network |
CN113255568B (en) * | 2021-06-15 | 2021-09-24 | 湖南星汉数智科技有限公司 | Bill image classification method and device, computer equipment and storage medium |
-
2015
- 2015-01-19 CN CN201510026025.3A patent/CN104573744B/en active Active
Non-Patent Citations (4)
Title |
---|
"Bird Species Categorization Using Pose Normalized Deep Convolutional Nets";Steve Branson et al.;《Eprint Arxiv》;20141231;第1页、图1-3 * |
"ImageNet classification with deep convolutional neural networks";A Krizhevsky et al.;《International Conference on Neural Information Processing Systems》;20121231;第1-9页 * |
"车型识别技术研究与应用";张颖;《中国优秀硕士学位论文全文数据库·信息科技辑》;20140215;I138-789 * |
王文."细粒度表情分类研究与设计".《中国优秀硕士学位论文全文数据库•信息科技辑》.2014,第23-52页. * |
Also Published As
Publication number | Publication date |
---|---|
CN104573744A (en) | 2015-04-29 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104573744B (en) | Fine granulation classification identifies and the part of object positions and feature extracting method | |
Yang et al. | Real-time face detection based on YOLO | |
Sirmacek et al. | A probabilistic framework to detect buildings in aerial and satellite images | |
CN106682598B (en) | Multi-pose face feature point detection method based on cascade regression | |
CN108549870B (en) | Method and device for identifying article display | |
CN106408030B (en) | SAR image classification method based on middle layer semantic attribute and convolutional neural networks | |
Zawbaa et al. | Automatic fruit classification using random forest algorithm | |
US9031317B2 (en) | Method and apparatus for improved training of object detecting system | |
US20110142335A1 (en) | Image Comparison System and Method | |
CN105488809A (en) | Indoor scene meaning segmentation method based on RGBD descriptor | |
CN110866079A (en) | Intelligent scenic spot real scene semantic map generating and auxiliary positioning method | |
CN105488536A (en) | Agricultural pest image recognition method based on multi-feature deep learning technology | |
CN109146889A (en) | A kind of field boundary extracting method based on high-resolution remote sensing image | |
CN108537102A (en) | High Resolution SAR image classification method based on sparse features and condition random field | |
CN108830312B (en) | Integrated learning method based on sample adaptive expansion | |
Peng et al. | Object-based change detection from satellite imagery by segmentation optimization and multi-features fusion | |
CN108509925B (en) | Pedestrian re-identification method based on visual bag-of-words model | |
Yuan et al. | Learning to count buildings in diverse aerial scenes | |
CN105718552A (en) | Clothing freehand sketch based clothing image retrieval method | |
CN104809471B (en) | A kind of high spectrum image residual error integrated classification method based on spatial spectral information | |
CN111368682A (en) | Method and system for detecting and identifying station caption based on faster RCNN | |
CN106127733A (en) | The method and apparatus of human body target identification | |
CN103065163A (en) | Rapid target detection and recognition system and method based on static picture | |
CN115147363A (en) | Image defect detection and classification method and system based on deep learning algorithm | |
CN102930291A (en) | Automatic K adjacent local search heredity clustering method for graphic image |
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 |