CN105894030B - High-resolution remote sensing image scene classification method based on layering multiple features fusion - Google Patents
High-resolution remote sensing image scene classification method based on layering multiple features fusion Download PDFInfo
- Publication number
- CN105894030B CN105894030B CN201610201638.0A CN201610201638A CN105894030B CN 105894030 B CN105894030 B CN 105894030B CN 201610201638 A CN201610201638 A CN 201610201638A CN 105894030 B CN105894030 B CN 105894030B
- Authority
- CN
- China
- Prior art keywords
- cluster
- feature
- remote sensing
- image
- training
- 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.)
- Expired - Fee Related
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
-
- 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
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a kind of high-resolution remote sensing image scene classification methods based on layering multiple features fusion.The present invention carries out clustering to training sample image set according to combination global characteristics, training sample image set is divided into and is suitable for the subset using global characteristics expression and the subset suitable for being expressed with local feature, global characteristics classifier then is respectively trained using two subsets and feature classifiers are simplified in part；It is selected from two classifiers in test according to cluster analysis result and is best suited for the classifier of test sample to classify to test sample.The present invention has high nicety of grading, experiments verify that, population mean nicety of grading of the invention can reach 96%, be higher than existing typical classification method.
Description
Technical field
The present invention relates to Remote Sensing Image Processing Technology more particularly to a kind of high-resolution based on layering multiple features fusion are distant
Feel image scene classification method.
Background technique
In recent years, high-resolution earth observation technology is grown rapidly, in land investigation, urban planning and army
The numerous areas such as thing played an important role.Compared to middle low resolution remote sensing images, high-resolution remote sensing image can be mentioned
For detailed terrestrial object information, but the distribution of various atural object space structures is more complex.With the raising of remote sensing images resolution ratio, spatially
Object geological information, texture information etc. are more obvious, but also bring the spectral differences opposite sex enhancing of similar atural object, different atural objects simultaneously
Spectrum heterogeneity reduce the problems such as.Therefore, although high score remote sensing image provides more fine earth's surface and portrays, automation
The difficulty of task of information extraction greatly increases.In remote sensing images technology, either specialty information extraction, motion change prediction are gone back
It is that the foundation of Remote Sensing Database all be unable to do without classification, how develops the new method for being suitable for remote sensing image classification to improve point
Class precision is urgent problem.
Traditional remote sensing images visual interpretation method does not require nothing more than professional experiences abundant and sufficient field investigation
Data, and on the basis of being built upon specific priori knowledge due to this recognition methods, identification difficulty is larger, efficiency compared with
It is low.Research about remote sensing images scene classification technology has had already appeared many methods at present.Not according to iamge description mode
Together, present image classification method can substantially be divided into: the description method based on global characteristics and the description side based on local feature
Method.Common global characteristics include: color histogram, LBP (Local binary pattern), Gabor textural characteristics etc..Base
In local feature Image Description Methods since constant feature can be found in the ever-changing image of same category, by
Numerous researchers' favor.Document [Ji RR, Yao HX, Liu W.Task-dependent visual-codebook
Compression [J] .IEEE Transactions on Image Processing, 2012,21 (4): 2282-2293.] make
Replace sparse coding with hard allocated code, Bof feature is compressed to make up coding method using supervision dictionary learning algorithm
Brought deficiency.Bolovinou et al. [Bolovinou A, Pratikakis I, Perantonis S.Bag of
spatio-visual words for context inference in scene classification[J].Pattern
Recognition, 2013,46 (3): 1039-1053.] vision word based on ordered space structural relation is proposed, in feature
It joined contextual information in expression.Zheng et al. [Zheng Y, Lu H, Jin C, et al.Incorporating
spatial correlogram into bag-of-features model for scene categorization[C]//
Proceedings of the 9th Asian Conference on Computer Vision.Xi an,China:
Springer, 2010:333-342.] thought based on color correlogram, it proposes space correlation figure concept and is applied to view
Feel word packet feature representation.
Since remote sensing images terrestrial object information is extremely abundant, but various atural object spatial distributions are more complex, similar target present compared with
Big spectrum is heterogeneous, and the sorting technique based on pixel be easy to cause more mistake point, leakage point phenomenon, so as to cause nicety of grading compared with
It is low.For remote sensing images, single features can only describe its part attribute, lack enough differentiation information, in image type ratio
Preferable classification generally can not be obtained in the case where more close (such as residential areas of denseness, middle density residential areas, sparse residential quarter)
Effect, in addition, the global characteristics description of scene image and local feature description and each own advantage: global characteristics are description images
Global information, can reflect the overall structure of entire scene image；The part that local feature description identifies target area is special
Property there is good adaptability, can be very well each target of scene switching, including different perspectives, different illumination intensity situations such as
Ground adapts to.Be not quite similar for different target, the various characteristic validities in image, as a result, reasonable amalgamation of global characteristics and
Local feature can improve the precision of remote sensing images scene classification to a certain extent.How correctly to merge different characteristic is figure
As a current research hotspot of classification field.Zhang Lei et al. [Zhang Lei, the EO-1 hyperion that the peak Shao Zhen improved OIF and SVM are combined
Classification of remote-sensing images [J] Surveying and mapping, 2014,39 (011): 114-117.] propose a kind of improved optimum index of combination
(OIF) and support vector machines carries out remote sensing image classification, and this method can effectively extract optimal bands combined, and obtain compared with high-class
Precision.For tradition fusion, spatially and spectrally characterization method only uses single space characteristics, and underuses double high-resolution
Feature, Wang Zengmao et al. proposition textural characteristics and morphological feature are merged, then carry out image classification [Wang Zengmao, Du Bo,
Zhang Liangpei, Hyperspectral Image Classification method [J] the photonics journal for waiting to merge based on textural characteristics and morphological feature, 2014,43
(8):810002-0810002.].For the terrestrial object information for making full use of remote sensing images, [Wang Yunyan, He Chu are applied Wang Yunyan et al.
Peak waits the blending algorithm of feature selecting bilayer SVM to be used for Classification of Polarimetric SAR Image [J] Wuhan University Journal information science version,
2015,40 (9): 1157-1162.] Feature Fusion Algorithm of feature selecting bilayer support vector machines is proposed, make full use of feature
Between completeness and heterogeneite, form the combination of more effective feature for image classification.Lift a sail et al. [it lifts a sail, Du Bo, Zhang Liang
Training, waits hyperspectral image classification method [J] computer science of a kind of combination wave band grouping feature of and morphological feature,
2014,41 (12): 275-279.] propose the hyperspectral image classification method based on band grouping feature and morphological feature, knot
It closes spatially and spectrally feature and improves nicety of grading.[Xiong Yu, Zuo little Qing, Huang Liang wait based on the coloured silk of combination of multiple features to Xiong Yu et al.
Color remote sensing image classification studies [J] laser technology, 2014,38 (2): 165-171.] it proposes based on color and texture-combined
Colouring information and texture information are combined as feature vector, are carried out to remote sensing images by Classifying Method in Remote Sensing Image, this method
Classification, and being experimentally confirmed this method is a kind of effective image classification method.Sheng et al. [Sheng G, Yang W,
Xu T,et al.High-resolution satellite scene classification using a sparse
coding based multiple feature combination[J].International journal of remote
Sensing, 2012,33 (8): 2395-2412.] it is distant for high-resolution using the method based on multiple features fusion sparse coding
Feel image classification, the probability in single feature channel is together in series as final fusion results.Fernando et al.
[Fernando B,Fromont E,Muselet D,et al.Discriminative feature fusion for image
classification[C]//Computer Vision and Pattern Recognition(CVPR),2012IEEE
Conference on.IEEE, 2012:3434-3441.] fusion method of logic-based recurrence is proposed, it is set using LR output
Count a kind of learning method of new SVM classifier efficient frontier core.Shao et al. [Shao W, Yang W, Xia G S, et
al.A hierarchical scheme of multiple feature fusion for high-resolution
satellite scene categorization[M]//Computer Vision Systems.Springer Berlin
Heidelberg, 2013:324-333.] it proposes histogram intersection kernel support vectors machine is respectively adopted in the different classifications stage
(HIK-SVM) with L1 normal form regularization logistic regression classifier (L1R-RLC), by the soft general of the obtained different characteristic of HIK-SVM
Rate, which is merged and imported in L1R-LRC, obtains final probability.
Although the above-mentioned prior art respectively has a feature, it is generally existing higher to the scene classification precision of certain particular categories and
The problem unsatisfactory to the nicety of grading of other classification scenes.
Summary of the invention
Technical problem to be solved by the present invention lies in overcome only to divide partial category scene present in the prior art
Class effect is preferably insufficient, provides a kind of high-resolution remote sensing image scene classification method based on layering multiple features fusion, right
It can get high nicety of grading in scene of all categories.
The present invention specifically uses following technical scheme to solve above-mentioned technical problem:
A kind of high-resolution remote sensing image scene classification method based on layering multiple features fusion, including training stage and survey
The examination stage；The training stage the following steps are included:
Step 1, the combination global characteristics for extracting training sample image, and clustering is carried out to training sample image, it obtains
To optimum cluster result；The combination global characteristics are composed of at least two global characteristics；
Step 2 selects the class cluster that cluster coincident indicator is greater than a preset threshold from optimum cluster result, with selected
All training sample images in class cluster carry out the training of disaggregated model based on combination global characteristics as the first training subset,
Obtain global characteristics classifier；Using remaining training sample image as the second training subset, to training sample image therein
Local shape factor is carried out, and carries out feature using frequent item set mining and simplifies, obtained feature of simplifying is then based on and carries out
The training of disaggregated model obtains part and simplifies feature classifiers；The cluster coincident indicator of any sort cluster specifically: in such cluster
That the most quantity accounting of a kind of training sample image in all training sample images of such cluster of quantity；
The test phase the following steps are included:
Step 3 selects class cluster center and test sample image from the optimum cluster result and is combining global characteristics sky
Between the middle the smallest class cluster of distance, and judge such cluster whether be cluster coincident indicator be greater than the preset threshold class cluster, such as
It is that the combination global characteristics of test sample image are then inputted into the global characteristics classifier and are classified；It is surveyed if not, extracting
It tries the local feature of sample image and the frequent item set that is excavated using the training stage is carried out feature to it and simplified, tested
Sample image simplifies feature, and feature classifiers progress is then simplified in the feature input part of simplifying of test sample image
Classification.
Preferably, the method for the clustering is specific as follows: under inhomogeneity number of clusters respectively to training sample image into
Row cluster, and premium class number of clusters is therefrom selected according to DB Cluster Validity Index or improved DB Cluster Validity Index, it is optimal
Cluster result corresponding to class number of clusters is optimum cluster result.
Preferably, the combination global characteristics are by including that at least two global characteristics of color characteristic and textural characteristics combine
It forms.Wherein, the color histogram feature of the preferred HSL color space of the color characteristic, the preferred Gabor line of textural characteristics
Manage feature.
Preferably, the local feature is SIFT feature.
Preferably, when simplifying using frequent item set mining progress feature, frequent two item collection in transaction library is only excavated.
Preferably, the value range of the threshold value is [0.9,1].
Preferably, for train global characteristics classifier and part simplify feature classifiers disaggregated model be support to
Amount machine.
Compared with prior art, the invention has the following advantages:
The present invention carries out clustering to training sample image set according to combination global characteristics, by training sample image collection
It closes to be divided into and is suitable for the subset using global characteristics expression and the subset suitable for being expressed with local feature, then utilize two subsets point
Not Xun Lian global characteristics classifier and part simplify feature classifiers；In test according to cluster analysis result from two classifiers
In select and be best suited for the classifier of test sample to classify to test sample.The present invention has high nicety of grading,
Experiments verify that population mean nicety of grading of the invention can reach 96%, it is higher than existing typical classification method.
Detailed description of the invention
Fig. 1 is residential areas of denseness, middle density residential areas, sparse residential quarter classification schematic diagram；
Fig. 2 is the training process in the specific embodiment of the invention；
Fig. 3 is the test process in the specific embodiment of the invention；
Fig. 4 is the method for the present invention figure compared with each single features are for classification results of all categories；
Fig. 5 is the method for the present invention nicety of grading confusion matrix figure of all categories；
Fig. 6 is the method for the present invention figure compared with the population mean nicety of grading of the prior art.
Specific embodiment
Technical solution of the present invention is described in detail with reference to the accompanying drawing:
For remote sensing images, single features can only describe its part attribute, lack enough differentiation information, in image class
Type leads in the case where more close (as shown in figure 1 from left to right shown in residential areas of denseness, middle density residential areas, sparse residential quarter)
Cannot often obtain preferable classifying quality, and due in remote sensing images some classification color of image features it is more apparent and some
Classification image local feature is more apparent, and it is lower to will lead to nicety of grading if only with single features, in addition, being based on global characteristics
Advantage is respectively had its own again with local feature description；Global characteristics are the global informations for describing image, can reflect entire scene figure
The overall structure of picture；Local feature description is to each target of scene switching, including situations such as different perspectives, different illumination intensity
It can be well adapted for.Reasonable amalgamation of global characteristics and local feature as a result, can improve remote sensing images scene to a certain extent
The precision of classification.
Based on the above thinking, the present invention is first depending on combination global characteristics and carries out cluster point to training sample image set
Training sample image set is divided into and is suitable for the subset using global characteristics expression and the son suitable for being expressed with local feature by analysis
Then collection is respectively trained global characteristics classifier using two subsets and feature classifiers is simplified in part；In test according to poly-
Alanysis result is selected from two classifiers is best suited for the classifier of test sample to classify to test sample.
For the ease of public understanding, to carry out technical solution of the present invention with a preferred embodiment of the present invention below detailed
Explanation.Training stage in the present embodiment the following steps are included:
Step 1, the combination global characteristics for extracting training sample image, and clustering is carried out to training sample image, it obtains
To optimum cluster result；The combination global characteristics are composed of at least two global characteristics；
Firstly the need of the extraction for carrying out global characteristics to the training sample in training sample image set, the training sample figure
Image set close in include each classification remote sensing images, known to the classification of every width training sample image.Training sample image set can
It is obtained from existing remote sensing image data library.
In order to overcome the limitation of single features, combination global characteristics that the present invention is combined into using a variety of global characteristics into
Row clustering.As being spoken of in background technique, common image overall feature includes color characteristic, LBP feature, texture
Feature etc..Wherein, the color characteristic of image is smaller to the dependence at the size of image itself, direction, visual angle, and common color is straight
Square figure feature describes different color ratio shared in entire image.Texture is a kind of important space letter of remote sensing images
Breath, with the raising of resolution ratio, the internal structure of atural object is more and more clear, this shows as the texture knot of atural object in remote sensing images
Structure is more and more obvious.Relative to spectral information, textural characteristics can reflect the well-regulated spatial variations of pixel in Target scalar.Cause
This, present invention preferably employs color characteristics and textural characteristics to be combined into combination global characteristics.
Color characteristic in the present embodiment is using based on HSL (Hue, Saturation, Lightness) color space
Color histogram feature is extracted, compared to rgb color space, the former more meets the vision perception characteristic of human eye.It is specifically extracted
Method is as follows:
For each pixel of image, its H, S, L * component are quantified as 8 equal portions respectively first, such HSL histogram is just
There are 512 intervals (bin), then count the histogram of all pixels, finally it is normalized using L1 normal form, to eliminate
The influence of image difference breadth size.
There is stronger texture description ability in view of Gabor filter, utilize Gabor filter extraction line in the present embodiment
Feature is managed, extracting method is as follows:
Using 6 directions, the Gabor filter of 5 scales carries out Gabor transformation to original image, forms 30 width images
Texture representation, then the mean value by calculating this 30 width image and variance are to obtaining the feature vector of 60 dimensions.
fMPEG7HT=[μ11,σ11,μ12,σ12,...,μ1S,σ1S,..,μRS,σRS] (1)
Wherein, uRSAnd σRSIt is the corresponding filtering mean value of output filter and variance in the direction R and S scale respectively.
The combination that extracted color histogram feature and Gabor textural characteristics concatenation are obtained the present embodiment is global special
Sign.
Clustering is using classical K-means clustering algorithm, the key task of clustering in the present embodiment
Cluster result is quantitatively evaluated, determines an optimal class number of clusters mesh.Cluster Validity evaluation, which refers to, comments cluster result
Valence is to determine the division of most suitable specific set of data and judge whether acquired results are effective, correct, main contents packets
It is evaluated when the excellent degree of cluster result and unknown cluster numbers of more different clustering algorithms in the case of cluster numbers known to including same poly-
The quality of class algorithm cluster result under the conditions of different cluster numbers, the latter are suitable for choosing the preferable clustering number of data set.At this
The latter's operating method is chosen in embodiment, it is therefore desirable to Cluster Validity be carried out to the K-means cluster result under different K values and commented
Existing all kinds of Cluster Validity evaluation indexes can be used in valence, the present invention, such as RAND index, DB index, KL index, Hart refer to
Mark etc., present invention preferably employs DB (Davies-Bouldin) Cluster Validity Index (referring to document [Davies D.L.,
Bouldin D.W..A cluster separation measure.1979.IEEETrans.Pattern Anal.Machine
Intell.1 (4) .224-227]) or improved DB Cluster Validity Index (referring to Chinese invention patent CN101853304B).
Optimum clustering number and corresponding optimum cluster are determined in the present embodiment using DB Cluster Validity Index as a result, specifically, this
Process of cluster analysis is specific as follows in embodiment:
(1) under different K values, based on extracted combination global characteristics using K-means algorithm to training sample image
Training sample in set is clustered；It is found through lot of experiments, for high-resolution remote sensing image, optimum cluster
Number is typically distributed in a lesser interval range, therefore the variation zone of K value can be selected in [INT (0.9N), 2N] range
Between, to effectively reduce the consumption for calculating time and computing resource；Wherein, INT (*) is bracket function, and N is high-definition remote sensing
The scene type sum of image；It is 21 due to commonly using scene type number at present, the variation range of K value is 20 in the present embodiment
~42.
(2) all kinds of separating degree d are calculatedij=d (vi,vj), wherein d (vi,vj) indicate two cluster centre vi,vjBetween away from
From.
(3) within-cluster variance in all kinds of clusters is calculatedWherein NiIndicate cluster ciThe number of middle clustering object；
Calculate similarity Rij, Rij=(Si+Sj)/dij。
(4) DB Validity Index is calculated, and optimum clustering number and corresponding optimum cluster are determined according to DB Validity Index
As a result；
Definition:Wherein NcIndicate cluster number；
N is set in the present embodimentcRange is [20,42], calculates best DB value in this range, corresponding cluster numbers and poly-
Class result is best cluster result.
In general, good clustering should response data collection as far as possible immanent structure, keep sample in class most
May be similar, sample is as dissimilar as possible between class, is exactly to keep inter- object distance minimum and between class distance pole from the point of view of distance
Big cluster is optimum cluster.Therefore definition or more analysis of DB Validity Index is combined as can be seen that DB value is smaller, then is gathered
Class effect is better.
In the present embodiment, for the different values of K, choosing cluster numbers corresponding when DB value minimum is best cluster
Number.
Step 2 selects the class cluster that cluster coincident indicator is greater than a preset threshold from optimum cluster result, with selected
All training sample images in class cluster carry out the training of disaggregated model based on combination global characteristics as the first training subset,
Obtain global characteristics classifier；Using remaining training sample image as the second training subset, to training sample image therein
Local shape factor is carried out, and carries out feature using frequent item set mining and simplifies, obtained feature of simplifying is then based on and carries out
The training of disaggregated model obtains part and simplifies feature classifiers；The cluster coincident indicator of any sort cluster specifically: in such cluster
That the most quantity accounting of a kind of training sample image in all training sample images of such cluster of quantity；
The calculation method of cluster coincident indicator of the invention are as follows: the largest number of one kind image numbers in class cluster/such
All picture numbers that cluster is included.The cluster coincident indicator of clusters all kinds of in optimum cluster result and preset threshold value are made into ratio
The preferable class cluster of consistency is clustered compared with selecting；For clustering the preferable class cluster of consistency, all training samples for being included with it
Training set of the image as global characteristics classifier；The training sample image that the class cluster of other cluster consistency differences is included then is made
The training set of feature classifiers is simplified for part.
The local image characteristics such as SIFT, SURF, Daisy can be used in the local feature, then utilize frequent item set mining
It carries out feature and simplifies that (this is the prior art, and [Liu Li, Zhan Yinyin, Luo Yang wait Scale invariant features transform for details, reference can be made to document
Operator summarizes [J] Journal of Image and Graphics, 2013,18 (8): 885-892]).Image is extracted in the present embodiment first
Then SIFT feature carries out feature using frequent item set mining and simplifies, the specific steps are as follows:
1, the SIFT local feature for extracting each image first, the appearance in sub- Expressive Features region is described using SIFT, will
Local feature in training image clusters local feature point set using K-means clustering algorithm after extracting and describing, often
A cluster centre indicates a vision word, and all cluster centres form vision word code book；
2, it using the method distributed firmly, using arest neighbors technology, obtains and Feature Descriptor is apart from nearest vision word
And the corresponding word frequency of each vision word is counted, form the vision bag of words character representation of image；
3, local feature corresponding to each vision word for each image in every class carries out local neighborhood feature group
It closes.A square area, the square side length (feature directly proportional to feature point scale delimited centered on each characteristic area
Size of Neighborhood be not it is fixed, it is related to characteristic area scale, and adapt to the variation of object size), it is all to appear in this
Feature in square area is that the local neighborhood feature of current signature combines and the vision word that repeats is by one
Processing obtains the local neighborhood feature combination of each feature of each image, and the local neighborhood feature of each feature is combined
Regard a transaction as；
4, transaction all in the same category image is put together into the input as frequent item set mining, used
Apriori Frequent Itemsets Mining Algorithm excavates transaction data base, sets minimum support threshold value as M, M is preset big
In 1 integer, the present invention is preferably taken as 200, excavates frequent two item collection all in transaction library (because of data of the embodiment of the present invention
It is sparse to collect feature, if excavating to frequent three item collection, four item collections ... less, the shadow that will lead to the numerous item collection of partial category vision intermediate frequency
Ring final classification result) frequently occur and relatively have the word combination of identification；
5, the combination of local neighborhood feature corresponding to each feature in all kinds of middle each images is operated, and the office is retained
Contain the word of frequent two item collection in the combination of portion's neighborhood characteristics, and calculate the vision word histogram of each image, by histogram
Standardization.
Disaggregated model is respectively trained using two training sets, global characteristics classifier can be obtained and feature point is simplified in part
Class device, the disaggregated model as used in the present embodiment are support vector machines, can be denoted as SVM I, SVM II respectively.
Fig. 2 shows the training process in the present embodiment.
Test phase in the present embodiment is specific as follows:
Step 3 selects class cluster center and test sample image from the optimum cluster result and is combining global characteristics sky
Between the middle the smallest class cluster of distance, and judge such cluster whether be cluster coincident indicator be greater than the preset threshold class cluster, such as
It is that the combination global characteristics of test sample image are then inputted into the global characteristics classifier and are classified；It is surveyed if not, extracting
It tries the local feature of sample image and the frequent item set that is excavated using the training stage is carried out feature to it and simplified, tested
Sample image simplifies feature, and feature classifiers progress is then simplified in the feature input part of simplifying of test sample image
Classification.
The color histogram feature of the HSL color space of test sample is extracted first, in accordance with mode identical with the training stage
And Gabor textural characteristics, and concatenated as combination global characteristics；Then obtained most using the training stage to test sample
Excellent cluster result carries out rough sort: test sample is calculated separately at a distance from all kinds of cluster centers, if the class nearest with test sample
The affiliated class cluster in cluster center is the preferable class cluster of above-mentioned cluster consistency, then uses to the test sample and be based on color and textural characteristics
Classifier (i.e. global characteristics classifier) classify；Conversely, then extracting local feature to the test sample and based on training
The frequent item set that stage excavates carries out feature to it and simplifies, then using based on classifier (the i.e. part essence for simplifying feature
Simple feature classifiers) classify.Finally obtain the accurate scene type of the test sample.
Fig. 3 shows the test phase process in the present embodiment.
In order to verify the effect of the method for the present invention, following proving and comparisom experiment has been carried out.
The computer that inventive algorithm is realized is configured to i5 processor, 64 bit manipulation systems；Algorithm realizes that environment is
Matlab2010b and Visual C++6.0.
Experiment use image data set be large-scale remote sensing image classification test data set public in the world [Yang Yi,
Shawn N.Spatial pyramid co-occurrence for image classification[C]//IEEE
International Conference on Computer Vision.2011:1465-1472.], it covers Nature and Man work
The number of scene totally 21 kinds of classifications, every kind of classification image is 100, and the average-size size of all color images, image is
256*256 pixel, these images are downloaded at United States Geological Survey (United States Geological Survey),
Including scene have: farmland, aircraft, ball park, beach, building, jungle, residential areas of denseness, forest, highway,
Golf course, harbour, crossroad, middle density residential areas, prefabricated house, overline bridge, parking lot, river, runway, sparse house
Area, oil storage tank, tennis court.These images reflect the different scenes of every class scene image, and have good representativeness.Under
In the experiment of face, unless otherwise noted, 80 width images are extracted at random from every class as training sample, remaining 20 width image conduct
Test sample.
80 width images are extracted from each classification as training sample, utilize support vector machines training classifier, residue 20
Width image is as test sample.Each feature design parameter is provided that Color color histogram feature vector dimension is 512；
Gabor texture feature vector dimension is 60 dimensions；Based on local feature and carry out feature mining the dimension for simplifying feature be 1000
Dimension is utilized respectively these three single image features and the at present more popular spatial pyramid Matching Model based on sparse coding
(ScSPM) feature is compared with the method for the present invention.In addition, we are also and directly using Frequent Itemsets Mining Algorithm to SIFT spy
The method excavated is levied to be compared and (be denoted as " frequent item set mining " method in Fig. 4 and Fig. 6).
Classification results such as Fig. 4, it can be seen from the figure that color histogram feature is in golf course and class of harbours problem
In classification accuracy rate it is higher, this is because the color characteristic of these classification images is more apparent in entire image；Intensively living
It is showed in the classification problems such as quarter and crossroad very poor.
Nicety of grading of the Gabor characteristic in farmland and runway classification problem is compared with other classifications height, this is because these classes
Texture is more visible in other image；It is showed in ball park, building, oil storage tank classification problem very poor.
ScSPM feature shows preferably in ball park, beach, runway classification problem, this is because this kind of image has absolutely
To space arrangement, wherein in the class image of beach, in spite of the bad discrimination in coastline in some particular orientations, but can be with
It is also such for runway class image by the sandy beach or the spray of same image-region come resolution image classification；For building
Object, the performance of oil storage tank classification problem are slightly poor.
Mining Frequent Itemsets are for building, residential areas of denseness, middle density residential areas, sparse residential quarter classification problem
In behave oneself best, but farmland classification problem show it is poor.
Although each feature has it to show preferable several classifications, population mean nicety of grading is not high, no
It can satisfy the requirement of nicety of grading in practice.
The classification results and the knot after two sub- Multiple Classifier Fusions that table 1 gives in the present invention two sub-classifiers
Fruit.
The classification results of 1 two sub-classifiers of table
Fig. 5 gives exhaustive division of all categories in present invention experiment as a result, as can be seen from the figure the method for the present invention is removed
Farmland, golf course, river classification image classification result are slightly less than outside single features, the nicety of grading of other classification images
It is above maximum nicety of grading when being classified using single features.The method of the present invention is significantly improved using single features
The frequent item set mining problem lower for oil storage tank classification accuracy rate, for beach, viaduct, parking lot and tennis court classification
Image classification performance has reached 100%, this has been absolutely proved for the different target in image, various characteristic validities phase not to the utmost
Together, there is complementarity to each other.
Fig. 6 gives single features and five empirical average nicety of grading contrast properties curve graphs of the invention, can from figure
To find out, the population mean nicety of grading highest of the method for the present invention reaches 96%.Good nicety of grading demonstrates calculation of the present invention
Method has more robustness.
Table 2 gives the method for the present invention and the nicety of grading of following several frequently seen multiple features fusion method compares:
Sheng et al. [Sheng G, Yang W, Xu T, et al.High-resolution satellite scene
classification using a sparse coding based multiple feature combination[J]
.International journal of remote sensing, 2012,33 (8): 2395-2412.] using based on multiple features
The method of sparse coding is merged for high-resolution remote sensing image classification (abbreviation SCMF), the probability in single feature channel is connected
As final fusion results.
Fernando et al. [Fernando B, Fromont E, Muselet D, et al.Discriminative
feature fusion for image classification[C]//Computer Vision and Pattern
Recognition (CVPR), 2012IEEE Conference on.IEEE, 2012:3434-3441.] propose logic-based
The fusion method (referred to as LRFF) of recurrence, this method generate visual dictionary to each feature first, then use logistic regression side
Method is inferred in certain kinds the most vision word of judgement index from multiple visual dictionaries, finally designs one kind using LR output
The learning method of new SVM classifier efficient frontier core.
Shao et al. [Shao W, Yang W, Xia G S, et al.A hierarchical scheme of
multiple feature fusion for high-resolution satellite scene categorization
[M] //Computer Vision Systems.Springer Berlin Heidelberg, 2013:324-333.] it proposes
Histogram intersection kernel support vectors machine (HIK-SVM) and L1 normal form regularization logistic regression point is respectively adopted in the different classifications stage
Class device (L1R-RLC), the soft probability of the obtained different characteristic of HIK-SVM is merged and import in L1R-LRC obtain it is final
Probability.
The method of the present invention is significantly better than SCMF and LRFF method as can be seen from Table 2, and than HIK-SVM, the side L1R-RLC
Method is high by 3.23%, illustrates that the method for the present invention is a kind of more effective multiple features fusion method.
2 the method for the present invention of table and various existing multiple features fusion classification accuracy comparisons
Claims (9)
1. a kind of high-resolution remote sensing image scene classification method based on layering multiple features fusion, including training stage and test
Stage；It is characterized in that, the training stage the following steps are included:
Step 1, the combination global characteristics for extracting training sample image, and clustering is carried out to training sample image, it obtains most
Excellent cluster result；The combination global characteristics are composed of at least two global characteristics；
Step 2 selects the class cluster that cluster coincident indicator is greater than a preset threshold from optimum cluster result, with selected class cluster
In all training sample images as the first training subset, the training of disaggregated model is carried out based on combination global characteristics, is obtained
Global characteristics classifier；Using remaining training sample image as the second training subset, training sample image therein is carried out
Local shape factor, and carry out feature using frequent item set mining and simplify, it is then based on obtained feature of simplifying and classifies
The training of model obtains part and simplifies feature classifiers；The cluster coincident indicator of any sort cluster specifically: quantity in such cluster
That the most quantity accounting of a kind of training sample image in all training sample images of such cluster；
The test phase the following steps are included:
Step 3 selects class cluster center with test sample image in combining global characteristics space from the optimum cluster result
Apart from the smallest class cluster, and judge whether such cluster is the class cluster for clustering coincident indicator and being greater than the preset threshold, if so, then
The combination global characteristics of test sample image are inputted the global characteristics classifier to classify；If not, extracting test specimens
The local feature of this image and the frequent item set excavated using the training stage are carried out feature to it and simplified, and test sample is obtained
Image simplifies feature, then inputs the part and simplifies feature classifiers and divided the feature of simplifying of test sample image
Class.
2. high-resolution remote sensing image scene classification method as described in claim 1, which is characterized in that the side of the clustering
Method is specific as follows: training sample image is clustered respectively under inhomogeneity number of clusters, and according to DB Cluster Validity Index or
Improved DB Cluster Validity Index therefrom selects premium class number of clusters, and cluster result corresponding to premium class number of clusters is optimal poly-
Class result.
3. high-resolution remote sensing image scene classification method as claimed in claim 2, which is characterized in that the inhomogeneity number of clusters
Value range is [INT (0.9N), 2N]；Wherein, INT (*) is bracket function, and N is the scene type of high-resolution remote sensing image
Sum.
4. high-resolution remote sensing image scene classification method as claimed in claim 2, which is characterized in that clustered using K-means
Method respectively clusters training sample image under inhomogeneity number of clusters.
5. high-resolution remote sensing image scene classification method as described in claim 1, which is characterized in that the combination global characteristics
It is composed of at least two global characteristics for including color characteristic and textural characteristics.
6. high-resolution remote sensing image scene classification method as claimed in claim 5, which is characterized in that the color characteristic is
The color histogram feature of HSL color space, the textural characteristics are Gabor textural characteristics.
7. high-resolution remote sensing image scene classification method as described in claim 1, which is characterized in that the local feature is
SIFT feature.
8. high-resolution remote sensing image scene classification method as described in claim 1, which is characterized in that the value model of the threshold value
It encloses for [0.9,1].
9. high-resolution remote sensing image scene classification method as described in claim 1, which is characterized in that for training global characteristics
It is support vector machines that the disaggregated model of feature classifiers is simplified in classifier and part.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610201638.0A CN105894030B (en) | 2016-04-01 | 2016-04-01 | High-resolution remote sensing image scene classification method based on layering multiple features fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610201638.0A CN105894030B (en) | 2016-04-01 | 2016-04-01 | High-resolution remote sensing image scene classification method based on layering multiple features fusion |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105894030A CN105894030A (en) | 2016-08-24 |
| CN105894030B true CN105894030B (en) | 2019-02-19 |
Family
ID=57012000
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610201638.0A Expired - Fee Related CN105894030B (en) | 2016-04-01 | 2016-04-01 | High-resolution remote sensing image scene classification method based on layering multiple features fusion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105894030B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106778768A (en) * | 2016-11-22 | 2017-05-31 | 广西师范大学 | Image scene classification method based on multi-feature fusion |
| CN106960221A (en) * | 2017-03-14 | 2017-07-18 | 哈尔滨工业大学深圳研究生院 | A kind of hyperspectral image classification method merged based on spectral signature and space characteristics and system |
| CN108427913B (en) * | 2018-02-05 | 2021-12-28 | 中国地质大学（武汉） | Hyperspectral image classification method combining spectral, spatial and hierarchical structure information |
| CN111191617B (en) * | 2020-01-02 | 2022-02-01 | 武汉大学 | Remote sensing scene classification method based on hierarchical structure |
| CN112070178B (en) * | 2020-09-18 | 2023-10-27 | 北京金山云网络技术有限公司 | Method and device for determining image sequence sample set and computer equipment |
| CN112232402B (en) * | 2020-10-13 | 2022-07-08 | 燕山大学 | Method for realizing hyperspectral camouflage target recognition based on improved OIF and SVM |
| CN112800826A (en) * | 2020-12-15 | 2021-05-14 | 四川航天神坤科技有限公司 | City change monitoring method and system based on remote sensing image classification |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102622607A (en) * | 2012-02-24 | 2012-08-01 | 河海大学 | Remote sensing image classification method based on multi-feature fusion |
| 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 |
| CN104680173A (en) * | 2015-01-26 | 2015-06-03 | 河海大学 | Scene classification method for remote sensing images |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9152881B2 (en) * | 2012-09-13 | 2015-10-06 | Los Alamos National Security, Llc | Image fusion using sparse overcomplete feature dictionaries |
-
2016
- 2016-04-01 CN CN201610201638.0A patent/CN105894030B/en not_active Expired - Fee Related
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102622607A (en) * | 2012-02-24 | 2012-08-01 | 河海大学 | Remote sensing image classification method based on multi-feature fusion |
| 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 |
| CN104680173A (en) * | 2015-01-26 | 2015-06-03 | 河海大学 | Scene classification method for remote sensing images |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105894030A (en) | 2016-08-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105894030B (en) | High-resolution remote sensing image scene classification method based on layering multiple features fusion | |
| Tong et al. | Channel-attention-based DenseNet network for remote sensing image scene classification | |
| CN104680173B (en) | A kind of remote sensing images scene classification method | |
| CN102622607B (en) | Remote sensing image classification method based on multi-feature fusion | |
| Zhu et al. | Bag-of-visual-words scene classifier with local and global features for high spatial resolution remote sensing imagery | |
| Xia et al. | Structural high-resolution satellite image indexing | |
| CN102073748B (en) | Visual keyword based remote sensing image semantic searching method | |
| CN103942564B (en) | High-resolution remote sensing image scene classifying method based on unsupervised feature learning | |
| CN109657610A (en) | A kind of land use change survey detection method of high-resolution multi-source Remote Sensing Images | |
| CN106503739A (en) | The target in hyperspectral remotely sensed image svm classifier method and system of combined spectral and textural characteristics | |
| CN109063754B (en) | Remote sensing image multi-feature joint classification method based on OpenStreetMap | |
| CN112488050A (en) | Color and texture combined aerial image scene classification method and system | |
| CN110309780A (en) | High resolution image houseclearing based on BFD-IGA-SVM model quickly supervises identification | |
| CN108629287A (en) | A kind of remote sensing image terrain classification method | |
| CN104182763A (en) | Plant type identification system based on flower characteristics | |
| He et al. | Object-oriented mangrove species classification using hyperspectral data and 3-D Siamese residual network | |
| Chen et al. | Measuring the effectiveness of various features for thematic information extraction from very high resolution remote sensing imagery | |
| Jiang et al. | Focal-test-based spatial decision tree learning: A summary of results | |
| CN108805057A (en) | A kind of SAR image oil depot area detection method based on joint significance analysis | |
| CN109117739A (en) | One kind identifying projection properties extracting method based on neighborhood sample orientation | |
| CN105069459B (en) | One kind is directed to High Resolution SAR Images type of ground objects extracting method | |
| CN112733736A (en) | Class imbalance hyperspectral image classification method based on enhanced oversampling | |
| Al-Ghrairi et al. | Classification of satellite images based on color features using remote sensing | |
| CN115223017A (en) | Multi-scale feature fusion bridge detection method based on depth separable convolution | |
| CN116486248A (en) | Multi-source multi-mode remote sensing image sea-air interface target collaborative detection and identification method |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190219 Termination date: 20200401 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |