CN105930876A - Plant image set classification method based on reverse training - Google Patents

Plant image set classification method based on reverse training Download PDF

Info

Publication number
CN105930876A
CN105930876A CN201610317701.7A CN201610317701A CN105930876A CN 105930876 A CN105930876 A CN 105930876A CN 201610317701 A CN201610317701 A CN 201610317701A CN 105930876 A CN105930876 A CN 105930876A
Authority
CN
China
Prior art keywords
training
image
sample
plant
reverse
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.)
Pending
Application number
CN201610317701.7A
Other languages
Chinese (zh)
Inventor
杜吉祥
张宇卉
翟传敏
范文涛
王靖
刘海建
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Huaqiao University
Original Assignee
Huaqiao University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Huaqiao University filed Critical Huaqiao University
Priority to CN201610317701.7A priority Critical patent/CN105930876A/en
Publication of CN105930876A publication Critical patent/CN105930876A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • 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
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/10—Character recognition
    • G06V30/19—Recognition using electronic means
    • G06V30/192—Recognition using electronic means using simultaneous comparisons or correlations of the image signals with a plurality of references
    • G06V30/194—References adjustable by an adaptive method, e.g. learning

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Databases & Information Systems (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention provides a plant image set classification method based on reverse training, and the method comprises four steps: (1), the preprocessing of a plant digital image; (2), the clustering and dividing of a training set, wherein a K-mean clustering method is employed; (3), the training of a classifier, and the classification of a testing set and a mixed training set; (4), integration classification, wherein a classification label of the testing set is outputted. The method employs a reverse training method, is good in classification effect, and is good in intelligibility.

Description

A kind of plant image collection sorting technique based on reverse training
Technical field
The present invention relates to image set classification field, a kind of plant image collection sorting technique based on reverse training.
Background technology
Plant image Research on classifying method based on image set to as if: how to utilize disaggregated model correctly to identify and treat point The image set classification of class.Categorizing system application based on image set is mainly: the identification of face, the classification of video image, Ornamental plant identification, the identification etc. of Medicinal Plants.Over the years Chinese scholars in classification problem based on image set also Propose a lot of algorithms.Therefore, the research of plant image categorizing system based on image set has major and immediate significance, and one Denier research success also puts into application, will produce huge Social and economic benef@.
At present, the whole world has welcome big data age, shows according to up-to-date investigation, within 2015, will have more than 200 Hundred million terminal units are connected on the Internet, and by these intelligent terminals, the data total amount of generation will reach 40zb, entirely The quantity of ball server also will rapidly increase.Just quickly setting out towards digital times in the world, has arrived the year two thousand twenty, the data of storage Total amount will be bigger than 2010 50 times.A lot of experts and scholars think this data huge explosion just as a kind of novel oil, if Can well utilize, using can be as a kind of novel Asset Type.Traditional data are the most all expressed by numeral, And under big data age background, the data such as such as image text audio visual have all reacted the daily life of people from microcosmic The every aspect lived, thus reflect the economic form of entire society.If able to these data collections are got up to carry out deeply Research and excavation, it finds that the deep rule hidden and phenomenon in these data.
In traditional plant image identification problem, it is decent that the training of grader and test are all based on single or a small amount of figure This carrying out.But being as the arrival of big data age, the collection of data stores the widespread development of technology such as sharing, very All can obtain substantial amounts of plant image data in many application scenarios, these plant images can be that classification problem provides substantial amounts of instruction Practice and test sample.And owing to the feature performance under the states such as varying environment time-temperature of the similar plant also differs, Even if the features such as its Ye Hua of homophyletic plant also have larger difference, therefore sorting technique based on single image is the biggest by having Limitation.Classification based on plant image collection will well solve problem above.As a example by concrete plant leaf blade identification, from one Opening or can be partitioned into the leaf image under several different situations in multiple plant images, these images are naturally constituted Multiple image collections corresponding to multiple plant individuals.At cognitive phase, equally collect plant individual to be identified many Width leaf image set, thus replace individual leaf image in traditional method.Use these image collections of different plant, just Plant leaf blade identification system based on image collection can be designed.
At present, classification problem based on image set, is more to apply in recognition of face problem based on video, this type of The main direction of studying of problem is how to process the illumination attitude visual angle in video the problem such as to block, and makes full use of more Image and the abundant changing pattern information that includes, set up face identification system based on image collection.From current figure From the point of view of the development of image set Recognition Theory, recognizer based on image collection difference is mainly concentrated in entering each data acquisition system Row mathematical modeling the and how model modeled being carried out on similarity measurement.
In current image collection category theory, the method for subspace modeling is that most study uses the widest one Idea about modeling.Along with the further investigation of sub-space learning method, people attempt with subspace representation image collection, can be by data Problem concerning study transfer the problem concerning study on subspace to, make solution simpler effectively.The work with initiative is The mutual subspace method (Mutual Subspace Method, MSM) that O.Yamaguchi etc. propose.This method is directly often Individual image collection is modeled as linear subspaces, then by the similarity between each sub spaces of main folder angular metric, finally by Neighbour judges classification results.Owing to image covers the most apparent changing pattern of object, data sample itself might not divide Cloth, in a linear space, therefore has much according to the improvement of MSM.Innovatory algorithm the earliest is the constraint of the propositions such as Fukui Subspace algorithm (Constrained Mutual Subspace Method, CMSM) mutually, this algorithm owning image set Sample point projects on the linear subspaces of a more identification, to solve sample point not asking at linear subspaces Topic.T.K.Kim etc. propose a kind of dependency relation discriminant analysis method (Discriminant-analysis of Canonical Correlations, DCC), this method utilizes and is similar to linear discriminant analysis (Linear Discriminant Analysis, LDA) thought, according to minimizing in class dependency and maximizing dependency between class, in the hope of the projection of subspace Transformation matrix.The major limitation of this kind of algorithm be only data sample be modeled as one linear spatially, then pass through phase To more weak discriminant information (angle of linear space) tolerance similarity.
Along with the development of manifold learning, it was recognized that complex data is often distributed in a nonlinear manifold On, the method modeled image set by the thought of manifold learning is arisen at the historic moment.Fan and Yeung etc. are modeled in sample data In one nonlinear manifold, then use hierarchical clustering to go to excavate the local linear structure of manifold, each manifold is modeled Become the set of multiple approximately linear subspace, use related angle to remove the similarity of metric linear subspace, last similarity Result is determined by the method for comprehensive ballot.Hadid and Pietikainen uses the linear model of local to go to simulate non-thread equally Property manifold.They are locally linear embedding into (Locally Linear Embedding, LLE) logarithm first by manifold learning arithmetic According to carrying out dimensionality reduction, then utilize k means clustering algorithm to mark off different Clustering Model, use cluster centre to represent each class Cluster sample, then obtains the distance between image set pair by the distance of tolerance and comprehensive these sample points pair.Wang etc. carry Going out to calculate the framework (Manifold-Manifold Distance, MMD) of manifold spacing, its basic thought is: first figure Image set closes and is modeled as a nonlinear manifold, then embeds clustering algorithm by maximum linear and manifold is expressed as one group of local Linear model, the problem therefore calculating manifold distance is converted into the calculating partial model problem to spacing.Chen etc. use connection Close rarefaction representation the subspace in manifold is modeled, then by calculating the reconstruction error of rarefaction representation, calculate son empty Between pair between distance.Shao's often literary composition etc. proposes plant species machine identification method based on manifold spacing, is first extracted plant Then multiple samples of each class are portrayed as a nonlinear manifold by the characteristics of image of image, and therefore identification problem converts For the distance between the different manifold of tolerance.
Except the modeling method of linear/non-linear manifold, people attempt being modeled image collection by more method. H.Cevikalp proposes to use affine hull and convex closure (Affine/Convex Hull) modeling, and algorithm uses affine hull, convex closure collection Close sets theory image set is showed, then use the method for convex optimization try to achieve two bag nearest virtual point between away from From, represent two distances gathered with this.Rarefaction representation is added in convex closure modeling by Yiqun Hu, proposes rarefaction approximation recently The method (Sparse approximated nearest points, SANP) of adjoint point.Convex closure is added again on this by Meng The constraint of regularization, reducing the complexity of SANP, improves classifying quality.
Image set is modeled by Wang etc. by the method for statistics, uses the second-order statistic of image set sample point i.e. to assist Variance matrix describes image set, then makes the second-order statistic data being distributed in Riemann manifold be mapped to European by kernel function Spatially, the LDA of classics or partial least squares algorithm is finally used to classify.Propose again after Lu to use multistage statistic Describe image set, combine single order average, second order covariance matrix and the information of front Third order statistic, then use Multiple Kernel Learning Method calculate the distance between image set.
Arif Mahmood etc. use semi-supervised spectral clustering to classify image set, first each class are modeled PCA In space, then using semi-supervised hierarchical clustering to cluster all of sample point, label only rises when terminating cluster Effect, then defines the distance between image set according to the probability distribution of every apoplexy due to endogenous wind sample.
Algorithm above major applications, in recognition of face based on video, or relatively lacks in the classification of plant image collection Weary.
Summary of the invention
Present invention is primarily targeted at and overcome drawbacks described above of the prior art, propose a kind of based on planting of inversely training Object image collection sorting technique.The method can provide a kind of big test sample collection and big training sample in classifying for plant image The efficient grader of collection, it is achieved Fast Learning and high accuracy identify.
The present invention adopts the following technical scheme that
A kind of plant image collection sorting technique based on reverse training, for test set to be identified is classified, its Being characterised by, the plant digital picture of acquisition known class label is as training set in advance, and remaining step is as follows:
1) training set image and test set image to be identified are carried out pretreatment to extract feature；
2) sample set of training set is clustered respectively, then split into combined training collection and remaining training set；
3) combined training collection is trained two graders with test set to be identified；
4) by step 2) remaining training set input step 3) two graders, i.e. can obtain remaining each sample in training set The class label of this set, obtains remaining the sample that in training set sample set, the output label of sample is identical with test set label Number, the known class label of the training set that wherein proportion is most is required test set class label.
Preferably, in step 1) described in pretreatment include sample carrying out binaryzation, smoothing, split and standardize, And extract Gist feature and PHOG feature.
Preferably, in step 2) in, described cluster uses K mean algorithm.
Preferably, in step 2) in, described is split as purposiveness fractionation or purposiveness selection.
Preferably, in step 2) in, the picture number step 1 to be close or equal to of described combined training collection) described in The picture number of test set to be identified.
Preferably, described two graders use support vector machine.
From the above-mentioned description of this invention, compared with prior art, there is advantages that
One simple sorting algorithm is expanded to solve many classification problems by the reverse training of the present invention.Compared to by a pair One, two classification of one-to-many expand to many classification problems, and reverse training is more efficient, it is only necessary to train two graders.Real Testing result to also indicate that, compared to existing image set sorting technique, the method can provide a kind of in classifying for plant image Big test sample collection and the efficient grader of big training sample set, it is achieved Fast Learning and high accuracy identify.
Accompanying drawing explanation
Fig. 1 is the model of the inventive method, and in figure, training set and the test set image of input are to have entered pretreatment；Figure In, Training set is expressed as training set image set set；Query Image set is expressed as single test image set； Divided is that the fractionation of training set purpose obtains combined training collection；Train binary classifier is expressed as combined training collection Grader is trained with test set；Test x2 on trained Classifier represents to test in the grader trained and tears open Divide remaining training set X2, obtain the training set set with test set identical category；
Fig. 2 is the flow chart of the inventive method；
Fig. 3 is in the case of extracting training set image pattern different characteristic, the classification accuracy of the inventive method；
Fig. 4 is existing image set sorting algorithm classification accuracy in selected plant image storehouse.
Detailed description of the invention
Below by way of detailed description of the invention, the invention will be further described.
With reference to Fig. 1, Fig. 2, a kind of plant image collection sorting technique based on reverse training, for by test set to be identified Classify, it is assumed that the test set of this band identification isOutput class distinguishing label y.The method can provide a kind of for Big test sample collection and the efficient grader of big training sample set in plant image classification, it is achieved Fast Learning and high accuracy are known Not.The plant digital picture of acquisition known class label is as training set in advance, and this training set includes M image set: X1, X2,...,XM, the c image set Xc={ xt|yt=c:t=1,2 ..., Nc, comprise NcOpen similar picture, this image set label For yc∈[1,2,...,M].Remaining step is as follows:
1) training set image and test set image to be identified are carried out pretreatment, including binaryzation, smooth, split, advise Generalized, obtains the image that pretreatment obtains, and extracts suitable characteristics of image, such as Gist feature and PHOG feature.The present invention The method using two characteristic line combinations carries out plant image and inversely trains, and characteristic information linear combination function is: F=α F1+βF2, wherein 0≤α≤1,0≤β≤1 and alpha+beta=1, F1: be characterized collection vector 1, F2 is characterized collection vector 2.
Wherein, a kind of effective ways describing picture shape information are gradient orientation histogram (Histogram of Orientated Gradients, HOG), HOG feature, can be fine by extracting edge or the distribution of gradient of regional area Ground characterizes edge or the gradient-structure of regional area internal object, and then characterizes the shape of target.Bosch proposes tower-type gradient direction Rectangular histogram (Pyramid Histogram of Orientated Gradients, PHOG), PHOG is excellent relative to tradition HOG's Point, is the feature that different scale can be detected, and ability to express is higher.First edge image is carried out pyramid and divides by PHOG feature Layer, then on every layer extract HOG feature, finally the characteristic vector of each layer is coupled together represent PHOG characteristic vector.This Outward, the contextual information extracting image that GIST feature is brief and concise, the vision of analog quantity people extracts process.According to Oliva and The method that Torralba proposes, is divided into 4 × 4 by processing, by the Gabor filter group in 8 directions of 4 yardsticks, the image obtained Grid, say, that the Gist intrinsic dimensionality finally obtaining image is 4 × 4 × 32=512.
2) K mean algorithm is used the training set of each class of training set to be polymerized to K bunch respectively, as training set Xc= {X1,X2,...,XK, generate after cluster bunch choose the picture of some and be combined into combined training training set X1= {x1,x2,…,xj, X1The number j of middle image is identical or close with the picture number of test set to be identified, and this step is referred to as For the purpose of Sexual behavior mode or purposiveness split.The remaining part of training set is remaining training set It is to haveOpen the X of imagecThe image pattern that the image pattern of the fractionation of image set is chosen in other words conj.or perhaps, wherein(whole Number), X1ForRemaining training set X2For X2=X X1,
3) by combined training collection X1Two graders are trained with test set Y to be identified.I.e. one two grader of training C1.Training is at X1, Y does.In test set Y to be identified, all image tagged are+1, combined training collection X1In all of figure Image scale is designated as-1.By X1In comprise and the generic image of Y as exterior point.Further, it is contemplated that test phase X2In image be input to Two grader C1In, grader needs the disposal ability to uncertain data, and support vector machine (SVM) is solving line There is convex optimization theoretical background the most in the problem that property can be divided, have the optimizing algorithm of fixing set pattern, in the problem of linear separability On have bigger disturbance rejection, over-fitting problem can be solved.So the present invention selects linear SVM (Support Vector Machine with a l inear Kernel).Sample-the label of a given training set is to (x(t),y(t)),y(t)∈ {+1 ,-1}, grader C1Optimization problem be:In formula, w is for being Number vector；C > 0 is punishment parameter.
4) respectively by remaining training set X2In the image of each classification be input to the grader C that previous step trains1Middle knowledge Not, i.e. can get the class label of each residue training set sample set, obtain remaining the defeated of sample in training set sample set The number of samples that outgoing label is identical with test set label, the known class label of the training set that wherein proportion is most is required survey Examination collection class label.Because X2In the label of each class be known, so the X identical with test set Y classification to be identified2In The most classification of certain quantity be exactly y.Concrete, the picture of input and the generic picture of Y will be marked as+1, be designated asCalculateClass label normalized frequency histogram h, if hcFor at X2Middle c class is identified as+1 figure The percentage ratio of sheet, then
Export the class label y of test set Y to be identified.X2In to be identified as the maximum class of number of pictures of+1 be anticipated The class label of Y
One simple sorting algorithm is expanded to solve many classification problems by the reverse training of the present invention.Compared to by a pair One, two classification of one-to-many expand to many classification problems, and reverse training is more efficient, it is only necessary to train two graders.Real Testing result to also indicate that, compared to existing image set sorting technique, the method that the present invention proposes is more efficient.
Illustrate
The plant leaf blade data base set up from Chinese Academy of Sciences's Hefei mechanical intelligence computing laboratory, this data base contains 221 kinds Plant totally one ten thousand seven thousand several leaf images, and gather in different time, different plants, the most different blade figures As being affected by factors such as illumination, visual angle, deformation.Selecting 83 kinds of plants in research, picture sum, more than 30,000, divides at random Become test set and training set.Every leaf image is individual vanes, carries out pretreatment, and image resolution ratio is 30 × 30, examines Consider and need great amount of samples collection to experiment, so each image pattern collection contains super many 200 pictures in Shi Yan.Randomly draw image Sample part composition training sample set, another part composition test sample collection.
A. the result that plant leaf blade different characteristic is extracted
Setting progression as L=3, gradient direction is divided into 20 intervals, and PHOG descriptor is just by 3 gradient orientation histograms Characteristic vector order couples becomes 420 dimensions, the 512 dimension GIST features in 8 directions of 4 yardsticks.
B. the linear combination of characteristics of image
The commonly used feature of blade classification has color characteristic, textural characteristics, shape facility, local feature.Visible single Feature can not well characterize blade information, so using the method for two characteristic line combinations to carry out plant image herein Reverse train, characteristic information linear combination function is:
F=α F1+βF2, wherein 0≤α≤1,0≤β≤1 and alpha+beta=1
C. Ensemble classifier
Use the Ensemble classifier algorithm in the present invention that plant image collection is classified.By PHOG feature GIST Feature Fusion Test as plant leaf blade feature.Contrast method have mutual subspace method (Mutual Subspace Method, MSM), based on manifold-manifold distance frame method (Manifold-Manifold Distance, MMD), divide based on popular differentiation Analysis method (Manifold Discriminant Analysis, MDA), covariance diagnostic method (Covariance Discriminative Learning, CDL), method (Affine/Convex Hull based on convex closure and affine hull distance Based Image set Distance, AHISD/CHISD), the rarefaction nearest point methods (Sparse of approximation Approximated nearest points, SANP), closest approach algorithm (Regularized Nearest based on regularization Points,RNP).In test, the parameter of above method is both referred to the optimal value of correlative theses, Setup Experiments. and Fig. 3 is for carry In the case of taking training set image pattern different characteristic, the classification accuracy of the inventive method, Fig. 4 is that the classification of existing image set is calculated Method is classification accuracy in selected plant image storehouse.
Above are only the detailed description of the invention of the present invention, but the design concept of the present invention is not limited thereto, all utilize this Design carries out the change of unsubstantiality to the present invention, all should belong to the behavior invading scope.

Claims (6)

1. a plant image collection sorting technique based on reverse training, for being classified by test set to be identified, it is special Levying and be, the plant digital picture of acquisition known class label is as training set in advance, and remaining step is as follows:
1) training set image and test set image to be identified are carried out pretreatment to extract feature；
2) sample set of training set is clustered respectively, then split into combined training collection and remaining training set；
3) combined training collection is trained two graders with test set to be identified；
4) by step 2) remaining training set input step 3) two graders, i.e. can obtain remaining each sample in training set The class label of set, obtains the sample that in the sample set of residue training set, the output label of sample is identical with test set label Number, the known class label of the training set that wherein proportion is most is required test set class label.
A kind of plant image collection sorting technique based on reverse training, it is characterised in that: in step 1) pretreatment described in includes sample carrying out binaryzation, smoothing, split and standardize, and extracts Gist feature and PHOG spy Levy.
A kind of plant image collection sorting technique based on reverse training, it is characterised in that: in step 2), in, described cluster uses K mean algorithm.
A kind of plant image collection sorting technique based on reverse training, it is characterised in that: in step 2), in, described is split as purposiveness fractionation or purposiveness selection.
A kind of plant image collection sorting technique based on reverse training, it is characterised in that: in step 2) in, the picture number step 1 to be close or equal to of described combined training collection) described in the picture number of test set to be identified Mesh.
A kind of plant image collection sorting technique based on reverse training, it is characterised in that: described two Grader uses support vector machine.
CN201610317701.7A 2016-05-13 2016-05-13 Plant image set classification method based on reverse training Pending CN105930876A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201610317701.7A CN105930876A (en) 2016-05-13 2016-05-13 Plant image set classification method based on reverse training

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201610317701.7A CN105930876A (en) 2016-05-13 2016-05-13 Plant image set classification method based on reverse training

Publications (1)

Publication Number Publication Date
CN105930876A true CN105930876A (en) 2016-09-07

Family

ID=56835861

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201610317701.7A Pending CN105930876A (en) 2016-05-13 2016-05-13 Plant image set classification method based on reverse training

Country Status (1)

Country Link
CN (1) CN105930876A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110118749A (en) * 2019-06-06 2019-08-13 南京林业大学 A kind of garden stuff pesticide residue detection method based near infrared spectrum

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102567742A (en) * 2010-12-15 2012-07-11 中国科学院电子学研究所 Automatic classification method of support vector machine based on selection of self-adapting kernel function
CN104517120A (en) * 2014-12-04 2015-04-15 西安电子科技大学 Remote sensing image scene classifying method on basis of multichannel layering orthogonal matching
CN105139028A (en) * 2015-08-13 2015-12-09 西安电子科技大学 SAR image classification method based on hierarchical sparse filtering convolutional neural network

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102567742A (en) * 2010-12-15 2012-07-11 中国科学院电子学研究所 Automatic classification method of support vector machine based on selection of self-adapting kernel function
CN104517120A (en) * 2014-12-04 2015-04-15 西安电子科技大学 Remote sensing image scene classifying method on basis of multichannel layering orthogonal matching
CN105139028A (en) * 2015-08-13 2015-12-09 西安电子科技大学 SAR image classification method based on hierarchical sparse filtering convolutional neural network

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
YU-HUI ZHANG: "《Reverse Training for 》", 《ICIC 2015:ADVANCED INTELLIGENT COMPUTING THEORIES AND APPLICATIONS》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110118749A (en) * 2019-06-06 2019-08-13 南京林业大学 A kind of garden stuff pesticide residue detection method based near infrared spectrum
CN110118749B (en) * 2019-06-06 2021-06-22 南京林业大学 Fruit and vegetable pesticide residue detection method based on near infrared spectrum

Similar Documents

Publication Publication Date Title
CN107016405B (en) A kind of pest image classification method based on classification prediction convolutional neural networks
CN107256246B (en) printed fabric image retrieval method based on convolutional neural network
CN102622607B (en) Remote sensing image classification method based on multi-feature fusion
CN106919920A (en) Scene recognition method based on convolution feature and spatial vision bag of words
CN107480620B (en) Remote sensing image automatic target identification method based on heterogeneous feature fusion
CN106650804A (en) Facial sample cleaning method and system based on deep learning features
Karamizadeh et al. Methods of pornography detection
Xiong et al. MSN: Modality separation networks for RGB-D scene recognition
Vaduva et al. A latent analysis of earth surface dynamic evolution using change map time series
Feng et al. Beyond tag relevance: integrating visual attention model and multi-instance learning for tag saliency ranking
Li et al. High-resolution SAR change detection based on ROI and SPP net
CN109117739A (en) One kind identifying projection properties extracting method based on neighborhood sample orientation
Siraj et al. Flower image classification modeling using neural network
CN111352926B (en) Method, device, equipment and readable storage medium for data processing
CN105447100B (en) A kind of cloud atlas search method based on shape feature
Kumar et al. Herbal plants leaf image classification using deep learning models based on augmentation approach
CN103336830A (en) Image search method based on structure semantic histogram
CN106570514A (en) Automobile wheel hub classification method based on word bag model and support vector machine
CN105930876A (en) Plant image set classification method based on reverse training
Li et al. Incremental learning of infrared vehicle detection method based on SSD
Chu et al. Scene classification based on SIFT combined with GIST
CN112070009B (en) Convolutional neural network expression recognition method based on improved LBP operator
Hapsari Weather classification based on hybrid cloud image using principal component analysis (PCA) and linear discriminant analysis (LDA)
CN107392225A (en) Plants identification method based on ellipse Fourier descriptor and weighting rarefaction representation
CN112926670A (en) Garbage classification system and method based on transfer learning

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20160907

RJ01 Rejection of invention patent application after publication