CN103336801B - Remote sensing image retrieval method based on multiple features LSH index combination - Google Patents

Abstract

The invention discloses a kind of remote sensing image retrieval method based on multiple features LSH index combination, belong to remote Sensing Image Retrieval technical field.The LSH of one of index technology best for high-dimensional feature space is indexed and introduces remote Sensing Image Retrieval field by the present invention, can effectively solve dimension disaster and retrieve time-consuming problem, it is achieved that the quick-searching of remote sensing images in scale；Simultaneously, the present invention is directed to LSH index and propose the index Effective exponent (FDIVI) of a new index Validity Index feature based distinguishing ability, select distinguish the feature of target and background by the LSH on each feature space being indexed assessment, thus be effectively increased the accuracy of retrieval result.Compared to existing technology, the present invention can be realized more rapidly and accurately the retrieval of magnanimity remote sensing image data.

Description

Remote sensing image retrieval method based on multiple features LSH index combination

Technical field

The present invention relates to remote Sensing Image Retrieval technical field, particularly relate to a kind of based on multiple features LSH index combination distant Sense image search method.

Background technology

Over nearly 40 years, earth observation data volume sharply increases along with the development of remote sensing technology, and various information systeies are over the ground The requirement of thing identification technology is more and more higher.Carry out remote sensing images the most quickly and efficiently automatically to classify and retrieve and become urgent need One of problem solved.

In general image retrieval, searched targets is piece image or is included in piece image, and range of search is figure As other independent some images in storehouse.Different with general natural image retrieval, remote Sensing Image Retrieval has a characteristic that to be treated Searched targets is in the remote sensing images that an amplitude ratio is bigger, other similar purpose be dispersed in cover certain territorial scope a width or In several other remote sensing images, scattered target has certain visual similarity.Content-based Remote Sensing Image Retrieval System based on content carries Since going out, researcher has been achieved for many achievements.As Zhu etc. uses Gabor textural characteristics to retrieve aerial image [Zhu Bin, Marshall R, Hsinchun C.Creating a large-scale content-based airphoto image Digital library [J] .IEEE Trans.on Image Processing, 2000,9 (1): 163-167], Lu Lizhen etc. [Lu Lizhen, Liu Renyi, Liu Nan. the remote sensing image retrieval method [J] of a kind of Fusion of Color and textural characteristics. China's image graphics Journal, 2004,9 (3): 328-332] Gabor textural characteristics is merged in proposition and color characteristic carries out remote Sensing Image Retrieval, and use The linear weighted function of texture and color characteristic Euclidean distance carrys out measured similarity.Zeng Zhiming et al. [Zeng Zhiming, Li Feng, Fu Kun. wait one Plant the Texture Segmentation Algorithm [J] of large scale remote sensing images information retrieval based on contents. Wuhan University Journal: information science version, 2005,30 (12): 1080-1083] the co-occurrence matrix textural characteristics improved is utilized to carry out large scale remote Sensing Image Retrieval.Geng Li River etc. [Geng Lichuan, Wu Yundong, Geng Zexun, Wang Yonggang. remote Sensing Image Retrieval [J] based on artificial immune system. China's image figure Shape journal, 2010,01:155-160] propose a kind of remote Sensing Image Retrieval algorithm based on artificial immune system.Zhu Jiali etc. [Zhu Jiali, Li Shijin, Wan Dingsheng, Feng Jun. feature based selects and the remote Sensing Image Retrieval [J] of semi-supervised learning. China's image Figure journal, 2011.16 (8): 1474-1482] selection of a kind of feature based and the remote Sensing Image Retrieval of semi-supervised learning are proposed Method.

The method realizing remote sensing images similarity is a lot, and that most basic is sequential scan algorithm (Sequential Scan Algorithm, SSA).Also having the tree algorithm divided based on space, such as: R-tree, Kd-tree, SR-tree, these are calculated The result that method returns is the most more accurate, but they time efficiencies on High Dimensional Data Set are the highest.Document [Weber R, Schek H,Blott S.A quantitative analysis and performance study for similarity-search methods in high-dimensional spaces[C]//Proc.of the24th Intl.Conf.on Very Large Data Bases (VLDB) .1998:194-205] in point out, in dimension higher than after 10, the calculation divided based on space Method is on the contrary not as linear search.

In remote Sensing Image Retrieval based on content, quick-searching goes out a satisfactory similar subgraph image set, than flower The expense long period retrieves the image set more attractive complying fully with requirement.And to different types of searched targets, feature should This is different, selects to represent that the feature of searched targets content, to carry out image retrieval, can improve accessibility further Energy.

Summary of the invention

The technical problem to be solved is to overcome prior art not enough, it is provided that a kind of based on multiple features LSH rope Draw the remote sensing image retrieval method of combination, different characteristic collection is set up LSH index respectively, and according to indexing Validity Index certainly The employing optimal characteristics adapted to retrieves image, can be effectively improved the accuracy of retrieval result.

The present invention solves above-mentioned technical problem the most by the following technical solutions:

Remote sensing image retrieval method based on multiple features LSH index combination, comprises the following steps:

Step 1, remote sensing images carrying out piecemeal, each image block constitutes subgraph as a subimage, all image blocks Image set；Step 2, in the feature space that at least two is different, respectively with described subgraph image set for data set set up LSH index； Described feature space is color feature space or/and textural feature space；

Step 3, for given query sample, the LSH index being utilized respectively each feature space carries out LSH retrieval, obtains LSH in each feature space retrieves result；

Step 4, retrieve result according to the LSH in each feature space, calculate the index Validity Index of each feature space: face The index Validity Index FDIVI in color characteristic space_cAnd the index Validity Index FDIVI of textural feature space_tRespectively according to Below equation calculates:

${\mathrm{FDIVI}}_c=\frac{R_t\left(1\right)}{R_t\left(0\right)}$

${\mathrm{FDIVI}}_t=\frac{R_t\left(0\right)}{R_t\left(1\right)}$

Wherein, $R_t\left(i\right)=\frac{1}{\left|C_i\right|}\underset{x\∈C_i}{\mathrm{\Σ}}D(x,q),i=\mathrm{0,1},$

In formula, | C_i| represent bunch C_iSize；C₀Represent the mesh that the subimage in the Hash bucket hit by query sample is constituted Mark bunch；C₁Represent by the non-targeted bunch constituted with the subimage in the Hash bucket of query sample arest neighbors；(x q) represents sample x to D And the distance between query sample q, for color characteristic, (x q) uses rectangular histogram to hand over distance to D；For textural characteristics, and D (x, q) Use Euclidean distance；R_tI () represents the average distance between query sample and each bunch；

Step 5, the color characteristic that selection index Validity Index value is minimum respectively, textural characteristics are as optimum color characteristic With optimum textural characteristics, and determine optimum color characteristic and the binaryzation weight of optimum textural characteristics in accordance with the following methods: index Validity Index value is more than 1, then weight is set to 0；Otherwise, weight is set to 1；

Step 6, when two kinds of optimal characteristics exist the feature that weight is 0, only use weight be 1 feature space in LSH retrieval result is as final remote Sensing Image Retrieval result；When two kinds of optimal characteristics weights are 1, optimum to both All subimages that LSH retrieval result in feature space is comprised, by being weighted the similarity of two kinds of optimal characteristics Linear operation generates comprehensive similarity, and several have the subimage of maximum comprehensive similarity as finally before therefrom selecting Remote Sensing Image Retrieval result.

Preferably, generate comprehensive by the similarity of two kinds of optimal characteristics is weighted linear operation described in step 5 Close similarity, the most in accordance with the following methods:

Given query sample and described subgraph is calculated respectively in optimum color feature space and optimum textural feature space The variance of the spacing of each subimage in image set；

Then according to below equation determines optimum color characteristic and the comprehensive similarity weight w of optimum textural characteristics_c、w_t:

$w_c=\frac{{\mathrm{\σ}}_c}{{\mathrm{\σ}}_c+{\mathrm{\σ}}_t},$

$w_t=\frac{{\mathrm{\σ}}_t}{{\mathrm{\σ}}_c+{\mathrm{\σ}}_t},$

Wherein, σ_cAnd σ_tRepresent respectively in optimum color feature space and optimum textural feature space given query sample with Described subimage concentrates the variance of the spacing of each subimage；

Given comprehensive distance D between query sample and arbitrary subimage i is obtained finally according to following formula_i, comprehensive distance is more Little, show that comprehensive similarity is the highest:

D_i=w_cD_ci+w_tD_ti,

Wherein, D_ciAnd D_tiRepresent given query sample in optimum color feature space and optimum textural feature space respectively And the distance between subimage i.

Further, described therefrom select before several have the subimage of maximum comprehensive similarity as final remote sensing Image searching result, the most in accordance with the following methods:

Selected part subimage is concentrated from described subimage, and in optimum color feature space and optimum textural feature space Middle average T calculating this partial subgraph picture distance each other respectively_cAnd T_t；

Determine similarity threshold T according to the following formula:

T=w_cT_c+w_tT_t,

Wherein, w_cAnd w_tIt is respectively optimum color characteristic and the comprehensive similarity weight of optimum textural characteristics；

In all subimages that last LSH retrieval result from two kinds of optimal characteristics spaces is comprised, choose with given The comprehensive distance between the query sample subimage less than or equal to similarity threshold T, as final remote Sensing Image Retrieval result.

Compared to existing technology, the method have the advantages that

First, the LSH of one of index technology best for high-dimensional feature space is indexed and introduces remote Sensing Image Retrieval by the present invention Field, can effectively solve dimension disaster in scale and retrieve time-consuming problem, it is achieved that the quick-searching of remote sensing images；

Secondly, the present invention is directed to LSH index and propose a new index Validity Index feature based distinguishing ability Index Effective exponent (feature discriminativeness-based indexing validation index, FDIVI), select distinguish the feature of target and background by the LSH on each feature space being indexed assessment, thus effectively Improve the accuracy of retrieval result.

Accompanying drawing explanation

Fig. 1 is the flow chart of present invention remote sensing image retrieval method based on multiple features LSH index combination；

Fig. 2 (a) Fig. 2 (c) is followed successively by the remote sensing figure using the inventive method, sequential scan algorithm, expert manually to give Settlement place retrieval result in Xiang；

Fig. 3 (a) Fig. 3 (c) is followed successively by the remote sensing figure using the inventive method, sequential scan algorithm, expert manually to give Cultured In The Lake Pen retrieval result in Xiang；

Fig. 4 (a) Fig. 4 (c) is followed successively by the remote sensing figure using the inventive method, sequential scan algorithm, expert manually to give Retrieval result in forest land, the Maoshan Mountain in Xiang；

Fig. 5 (a) Fig. 5 (c) is followed successively by the remote sensing figure using the inventive method, sequential scan algorithm, expert manually to give Soil erosion retrieval result near Xuzhou in Xiang.

Detailed description of the invention

Below in conjunction with the accompanying drawings technical scheme is described in detail:

The thinking of the present invention is LSH index technology to introduce the automatic of magnanimity remote sensing image data inquire about and in retrieval, profit Solve the dimension disaster in remote Sensing Image Retrieval with LSH index and retrieve time-consuming problem, thus realizing the quick of remote sensing images Retrieval；And the index proposing a new index Validity Index feature based distinguishing ability for LSH index is effective Sex index (feature discriminativeness-based indexing validation index, FDIVI), passes through LSH on each feature space is indexed assessment and selects distinguish the feature of target and background, thus be effectively improved retrieval knot The accuracy of fruit.

For the ease of public understanding technical solution of the present invention, the most first LSH index technology is briefly introduced.

1998, Indyk and Motwani [Indyk P, Motawani R.Approximate Nearest Neighbors:Towards Removing the Curse of Dimensionality[C]//Proc.of the30th Annual ACM Symp.on Theory of Computing, 1998:p604-613] propose quick for nearest neighbor search problem The theoretical basis of sense position Hash (Locality Sensitive Hashing, LSH) algorithm.1999, Gionis etc. [Gionis A,Indyk P,Motwani R.Similarity search in high dimensions via hashing [C]//Proc.of the25th International Conference on Very Large Data Bases(VLDB) .1999] it improved and tested, using the way of Hash to solve the quick-searching problem of high dimensional data.2002, Tang Pretty China etc. [Tang Junhua, Yan Baoping. rapid image retrieval [J] based on LSH index. computer engineering and application, 2002,24: 20-21,63] LSH Index Algorithm is applied in Content-Based Image Retrieval system, its performance is better than traditional indexing means. 2007, Lv et al. [Qin Lv, William Josephson, Zhe Wang, Moses Charikar, Kai Li, Multi- Probe LSH:Efficient Indexing for High-Dimensional Similarity Search[C]// Proceeding-VLDB07Proceedings of the33rd international conference on Very Large data bases.2007] on the basis of LSH algorithm based on entropy, it is proposed that investigate LSH (Multi-Probe) more and calculate Method, more effectively utilizes every Hash table to reach to improve the purpose of retrieval quality.2009, Cai Heng etc. [Cai Heng, Li Zhoujun, Sun Jian, Li Yang. Chinese text quick-searching [J] based on LSH. computer science .2009,36 (8): 201-204,230] for The LSH algorithm of binary vector has been done further improvement by text retrieval.

LSH Index Algorithm is by building Hash table and search algorithm two parts form, it is assumed that current subimage feature composition Data set is P.L is Hash table number, and each Hash table comprises the hash function of M stochastic generation, and Hash table width is W.Structure Build hash table algorithm as follows:

1, every kind of feature to subimage carries out the conversion of Hamming (Hamming) space, P is mapped to Hamming spatially.

2, suitable parameter L, M, W are chosen.Initialize hash function race H={h₁,...,h_L, calculate feature in the index Position, the value of each hash function is determined by following formula:

$h_{a,b}=\frac{a*v+b}{W}---\left(1\right)$

Wherein, a and b obeys gaussian random distribution, and v is the characteristic vector of subimage, obtains after calculating $g\left(v\right)=(h_{a_1,b_1}\left(v\right),...,h_{a_M,b_M}\left(v\right)).$

3, utilize these hash functions, feature is stored in corresponding hash table.

Search algorithm is as follows:

1, query sample q is selected, and by the maps feature vectors of q to Hamming space.

2, the structure of similar developing algorithm Hash table, calculates cryptographic Hash g (q) of q, returns in the Hash bucket that g (q) points to Object.

3, Candidate Set is chosen.For LSH multi-dimensional indexing, owing to, in same hash function race, random function is independent With distribution, so regarding any one Hash table as a grader.And importance between Hash table is phase in indexing With, general ballot method just can calculate the synthesis result of many subseries.

The present invention proposes to be applied in remote Sensing Image Retrieval LSH Index Algorithm, and its basic thought is: carry out subimage Specific hash, in every Hash table, two subimages similar on higher dimensional space can fall into same on the biggest probability In individual Hash bucket.Regarding the set that subimage forms as a small-sized image library, system is near with query image by returning One group of image as patibhaga-nimitta, then searches for the target image the most similar to query image in the image set returned, reduces and use The time that family is retrieved every time.

Different types of searched targets in remote sensing images, the description to its content, the effectiveness of different characteristic also differs, If able to select the optimal feature representing searched targets content to carry out image retrieval, then can be greatly enhanced accessibility Energy.According to research LSH directoried data set classifying quality carried out in the past, it was found that such a conclusion: the k of inquiry target is near Neighbour is mainly distributed in target bucket, and only minority can be distributed in the Hash bucket of target bucket.Remote sensing images subimage amount with The quantity of large database is compared less mutually, and has redundancy between subimage, so the present invention considers when index selection With the non-targeted bucket that target bucket distance is 1.

In the multiclass feature describing remote sensing images, targets different in color feature remote sensing images, textural characteristics Can make up the deficiency that spatial information is distributed by color characteristic, it is former that index corresponding to both features meets that grader selects Then, not only there is High Defferential degree but also there is high complementary information.Therefore, the present invention uses this two category feature to combine to carry out remote sensing images Retrieval.

The present invention utilizes the characteristic vector of sample image, according to the representativeness of query sample in LSH index classification more than bunch The collection center representativeness to whole gathering, it is proposed that a kind of index Validity Index feature based mirror for LSH index Index Effective exponent (the feature discriminativeness-based indexing validation of other ability Index, FDIVI), its computing formula is specific as follows:

$R_t\left(i\right)=\frac{1}{\left|C_i\right|}\underset{x\∈C_i}{\mathrm{\Σ}}D(x,q),i=\mathrm{0,1}---\left(2\right)$

${\mathrm{FDIVI}}_c=\frac{R_t\left(1\right)}{R_t\left(0\right)}$

${\mathrm{FDIVI}}_t=\frac{R_t\left(0\right)}{R_t\left(1\right)}---\left(3\right)$

Wherein, in formula (2), | C_i| represent bunch C_iSize, C₀Represent the subgraph in the Hash bucket hit by query sample As the target bunch constituted, C₁Represent by the non-targeted bunch constituted with the subimage in the Hash bucket of query sample arest neighbors.Q represents Query sample, R_tI () represents the average distance between query point and each bunch.For color characteristic, (x q) uses rectangular histogram to D Distance is handed over to calculate；For textural characteristics, (x q) uses Euclidean distance to calculate to D.

In formula (3), FDIVI_cRepresent the desired value of color characteristic, FDIVI in feature space_tRepresent the finger of textural characteristics Scale value.When considering two kinds of color characteristics, FDIVI_cRefer to that target value is the least, illustrate on the color feature space of its correspondence, R_t And R (0)_t(1) value difference is away from the biggest, and LSH index is the best to the classifying quality of subimage, and therefore this feature is elected as optimum color Feature.When considering two kinds of textural characteristics, similar color characteristic, choose FDIVI_tIt is worth less feature special as optimum texture Levy.

I.e. can get technical scheme according to above-mentioned analysis, come skill of the present invention with a specific embodiment below Art scheme is further described.

The remote sensing image retrieval method based on multiple features LSH index combination of the present invention, as it is shown in figure 1, include following step Rapid:

Step 1, remote sensing images carrying out piecemeal, each image block constitutes subgraph as a subimage, all image blocks Image set.

The present embodiment uses overlap partition strategy conventional in remote Sensing Image Retrieval.When overlapping size takes minima 1*1 Pixel, the most overlapping between subimage after piecemeal, some Small object may be divided into several subimages dissimilar with sample, Retrieval result is worst；When overlapping size takes maximum 63*63, the most intensive between subimage after piecemeal, retrieval result is the most accurate, but time Empty complexity is the highest.Owing to present invention employs index strategy, therefore between subimage block, overlapping degree can be of a relatively high, real Remote sensing images are preferably divided in border the image block of series of partially overlapping 64*64 size, and between block, overlapping size is 60* 60。

Step 2, in the feature space that at least two is different, set up LSH rope with described subgraph image set for data set respectively Draw；Described feature space is color feature space or/and textural feature space.

In this embodiment, HSI color characteristic, Lab color characteristic, gray level co-occurrence matrixes it are extracted respectively (Gray level Co-Occurrence Matrix, GLCM) textural characteristics and Gabor textural characteristics are selected multiple In feature space, respectively each subimage is carried out feature extraction；Then in each feature space, LSH index is set up respectively: by height Dimensional feature is mapped to Hamming space, generates the hash function race of LSH, then every kind of spy of antithetical phrase image set according to default parameter Levy and carry out position sensing Hash hash, set up LSH index.In this specific embodiment, institute in Hash table number L, each Hash table Hash function quantity M comprised and the preferred value of Hash table width W be: L=20, M=5, W=4.

It is relatively time-consuming owing to view picture remote sensing images are extracted various features ratio, so off-line completes this process effectively save use of energy The time of family inquiry.Therefore the present invention is preferably by step 1, step 2 off-line execution.

Step 3, for given query sample, the LSH index being utilized respectively each feature space carries out LSH retrieval, obtains LSH in each feature space retrieves result.

For given query sample, utilizing the data set that LSH indexed search is close is prior art, and here is omitted.

Step 4, retrieve result according to the LSH in each feature space, calculate the index Validity Index of each feature space: face The index Validity Index FDIVI in color characteristic space_cAnd the index Validity Index FDIVI of textural feature space_tRespectively according to Below equation calculates:

${\mathrm{FDIVI}}_c=\frac{R_t\left(1\right)}{R_t\left(0\right)}$

${\mathrm{FDIVI}}_t=\frac{R_t\left(0\right)}{R_t\left(1\right)}$

Wherein, $R_t\left(i\right)=\frac{1}{\left|C_i\right|}\underset{x\∈C_i}{\mathrm{\Σ}}D(x,q),i=\mathrm{0,1},$

In formula, | C_i| represent bunch C_iSize；C₀Represent the mesh that the subimage in the Hash bucket hit by query sample is constituted Mark bunch；C₁Represent by the non-targeted bunch constituted with the subimage in the Hash bucket of query sample arest neighbors；(x q) represents sample x to D And the distance between query sample q, for color characteristic, D (x, q) use rectangular histogram hand over distance (see document [M.J.Swain, D.H.Ballard,Color indexing,International Journal of Computer Vision,7:1,11-32 (1991)])；For textural characteristics, (x q) uses Euclidean distance to D；R_tI () represents between query sample and each bunch average Distance.

Step 5, the color characteristic that selection index Validity Index value is minimum respectively, textural characteristics are as optimum color characteristic With optimum textural characteristics, and determine optimum color characteristic and the binaryzation weight of optimum textural characteristics.

When considering two or more color characteristic, FDIVI_cRefer to that target value is the least, the color in its correspondence is described On feature space, R_tAnd R (0)_t(1) value difference is away from the biggest, and LSH index is the best, therefore by this feature to the classifying quality of subimage Elect optimum color characteristic as.When considering two or more textural characteristics, similar color characteristic, choose FDIVI_tIt is worth less Feature as optimum textural characteristics.

When the FDIVI index of selected optimal characteristics is more than 1, for color characteristic, represent query sample and target bunch it Between average distance less than with arest neighbors non-targeted bunch between average distance, illustrate can not significantly distinguish in color space Target bunch and arest neighbors non-targeted bunch, be now set to 0 by the weight of color characteristic, be otherwise 1；For textural characteristics, then it represents that Average distance between query sample and target bunch is more than the average distance between arest neighbors non-targeted bunch, illustrates at texture empty Can not significantly distinguish target bunch and arest neighbors non-targeted bunch between, now the weight of textural characteristics is set to 0, be otherwise 1.

Step 6, when two kinds of optimal characteristics exist the feature that weight is 0, only use weight be 1 feature space in LSH retrieval result is as final remote Sensing Image Retrieval result；When two kinds of optimal characteristics weights are 1, optimum to both All subimages that LSH retrieval result in feature space is comprised, by being weighted the similarity of two kinds of optimal characteristics Linear operation generates comprehensive similarity, and several have the subimage of maximum comprehensive similarity as finally before therefrom selecting Remote Sensing Image Retrieval result.

This specific embodiment uses following methods determine the comprehensive similarity between subimage and given query sample:

Given query sample and described subgraph is calculated respectively in optimum color feature space and optimum textural feature space The variance of the spacing of each subimage in image set；

Then according to below equation determines optimum color characteristic and the comprehensive similarity weight w of optimum textural characteristics_c、w_t:

$w_c=\frac{{\mathrm{\σ}}_c}{{\mathrm{\σ}}_c+{\mathrm{\σ}}_t},$

$w_t=\frac{{\mathrm{\σ}}_t}{{\mathrm{\σ}}_c+{\mathrm{\σ}}_t},$

Wherein, σ_cAnd σ_tRepresent respectively in optimum color feature space and optimum textural feature space given query sample with Described subimage concentrates the variance of the spacing of each subimage；

Given comprehensive distance D between query sample and arbitrary subimage i is obtained finally according to following formula_i, comprehensive distance is more Little, show that comprehensive similarity is the highest:

D_i=w_cD_ci+w_tD_ti,

Wherein, D_ciAnd D_tiRepresent given query sample in optimum color feature space and optimum textural feature space respectively And the distance between subimage i.

So can retrieve in all subimages that result is comprised by the LSH from two kinds of optimal characteristics spaces, select The partial subgraph picture that comprehensive similarity is the highest, as final retrieval result.This specific embodiment specifically uses following methods: From described subimage concentrate selected part subimage (the present embodiment is from subimage concentrate randomly select 10% subimage), And calculate this partial subgraph picture distance each other respectively in optimum color feature space and optimum textural feature space Average, be designated as T respectively_cAnd T_t；Then according to the comprehensive similarity power of the optimum color characteristic obtained and optimum textural characteristics Weight w_cAnd w_t, determine similarity threshold T:T=w_cT_c+w_tT_t；Last LSH retrieval result from two kinds of optimal characteristics spaces is wrapped In all subimages contained, choose and give comprehensive distance D between query sample_iThe subimage of≤T, as final remote sensing Image searching result.

Below by the remote Sensing Image Retrieval of different ground mulching being verified the effectiveness of the inventive method, and make It is estimated with precision ratio based on area and recall ratio.

If Area (s) is all similar to the query sample region area retrieved in inquiry, Area (v) is in image But the region area that be not retrieved similar with query sample, Area (u) is that retrieve and that query sample is unrelated area surface Long-pending, then the recall ratio retrieved and precision ratio can be expressed as:

Precision ratio: $\mathrm{precision}=\frac{\mathrm{Area}\left(s\right)}{\mathrm{Area}\left(s\right)+\mathrm{Area}\left(u\right)}---\left(4\right)$

Recall ratio: $\mathrm{recall}=\frac{\mathrm{Area}\left(s\right)}{\mathrm{Area}\left(s\right)+\mathrm{Area}\left(v\right)}---\left(5\right)$

Taking into account retrieval precision and efficiency, use image subblock size 64*64, between sub-block, piecemeal overlap size is 60 pictures Element.To user's index building time, the retrieval time of index of reference and use sequential scan algorithm is made to carry out similarity measurement and need The time three wanted carries out Statistical Comparison, and corresponding time showing is in Table 1.

The remote sensing images LSH indexing means of the different ground mulching of table 1 contrasted with the sequential scanning method time

Picture number	LSH builds the time (s)	LSH query time (s)	Sequential query time (s)
				2	110.5	0.031	6.692
3	64.49	0.024	10.37
				4	26.38	0.031	8.456

5

12.23

0.02

4.508

Found out by table 1, when single inquiry is carried out for all images, the query time of LSH index and sequential scan time Compare and greatly reduce.

Fig. 2 (a)～Fig. 5 (c) is to Soil erosion region near settlement place, Cultured In The Lake Pen, forest land, the Maoshan Mountain and Xuzhou respectively Retrieval result, original image pixels is respectively 1024*768,1024*768,800*749,812*531, has scaling herein.Its In, Fig. 2 (a), Fig. 3 (a), Fig. 4 (a), Fig. 5 (a) they are the retrieval results using the inventive method to obtain, Fig. 2 (b), Fig. 3 (b), figure 4(b), Fig. 5 (b) be the retrieval result utilizing sequential scan algorithm to obtain, Fig. 2 (c), Fig. 3 (c), Fig. 4 (c), Fig. 5 (c) are experts The similar area that really should retrieve manually be given.

As can be seen from the figure, compared with sequential scan algorithm, the retrieval effectiveness utilizing feature combination method to obtain compares Satisfactory, the region of each target area false retrieval missing inspection is less.For Fig. 2 (a) Fig. 2 (c), it can be seen that the inventive method Retrieve the settlement place region, one piece of the lower right corner that sequential scanning method does not retrieve, and the picture that sequential index retrieves is right The region of side is wrong, utilizes feature combination method then it can be avoided that false retrieval；Equally, Fig. 3 (a) Fig. 3 (c) can be seen that profit By the zone errors that inventive process avoids below the picture retrieved by sequential scanning method；Can in Fig. 4 (a) Fig. 4 (c) Finding out that forest land, the Maoshan Mountain regional extent that the inventive method retrieval obtains is more than sequential index, recall ratio and precision ratio are higher；Fig. 5 A, in () Fig. 5 (c), the inventive method retrieves Soil erosion near the Xuzhou, one piece of the lower right corner that sequential index does not retrieve Region.

The sharpest edges of the inventive method are combination of multiple features, and it is by mesh different in color feature remote sensing images Mark, introduces textural characteristics and makes up the deficiency that spatial information distribution is described by color characteristic.

When two width subimages are dissimilar on color characteristic, further by analyzing textural characteristics between image vector away from From thus become similar, near that block Xuzhou, the lower right corner in that block settlement place region in the lower right corner and Fig. 5 (a) in Fig. 2 (a) Soil erosion region, they are not the most consistent with user's query sample on color characteristic, when using sequential scanning method Time, these regions can be missed, and can be retrieved by analyzing textural characteristics.Similar, when two width subimages are in face On color characteristic similar time, when use sequential scanning method, these regions can be mistakenly detected out, the present invention by analyze texture Feature increases the distance between image vector, thus reduces error detection.As left in the region below picture in Fig. 3 (a) and Fig. 5 (a) Side region part, they compare similar with user's query sample on color characteristic, can draw it by analyzing textural characteristics Be not target subimage, this region can be avoided to be mistakenly detected out.

Therefore, after introduced feature selects, the method using color and textural characteristics to combine can be effective to searched targets region Return, and need not manual intervention, significantly reduce the burden of user.

Table 2 is that the inventive method and sequential scanning method use recall ratio based on area and the knot of precision ratio assessment Really, it can be seen that the present invention based on multiple features LSH index combination remote sensing image retrieval method on recall ratio and precision ratio all Increase.

Table 2 accuracy result

From above-mentioned experimental result it can be seen that the index index utilizing the present invention to propose selects the feature being suitable for, and move State determines that the weight of feature is rational, and remote sensing image retrieval method based on multiple features LSH index combination can obtain preferably Retrieval effectiveness, and need not manual intervention, retrieval performance can be effectively improved.Experimental result is it is also shown that based on interior In the remote Sensing Image Retrieval held, being not that the feature kind used is the most, the performance of retrieval is the best.Therefore by index knot Really, selecting optimal characteristics by feature selection, recycling index index automatically determines the weight of feature, uses suitable combination side Method carries out image retrieval and is possible not only to improve retrieval precision, moreover it is possible to effectively reduces the amount of calculation in retrieving, improves retrieval Speed.

Claims (6)

1. remote sensing image retrieval method based on multiple features LSH index combination, it is characterised in that comprise the following steps:

Step 1, remote sensing images carrying out piecemeal, each image block constitutes subgraph image set as a subimage, all image blocks；

Step 2, in the feature space that at least two is different, respectively with described subgraph image set for data set set up LSH index；Institute Stating feature space is color feature space or/and textural feature space；

Step 3, for given query sample, the LSH index being utilized respectively each feature space carries out LSH retrieval, obtains each spy Levy the LSH in space and retrieve result；

Step 4, retrieve result according to the LSH in each feature space, calculate the index Validity Index of each feature space:

The index Validity Index FDIVI of color feature space_cAnd the index Validity Index of textural feature space

FDIVI_tCalculate according to below equation respectively:

${\mathrm{FDIVI}}_c=\frac{R_t\left(1\right)}{R_t\left(0\right)}$

${\mathrm{FDIVI}}_t=\frac{R_t\left(0\right)}{R_t\left(1\right)}$

Wherein, $R_t\left(i\right)=\frac{1}{\left|C_i\right|}\underset{x\∈C_i}{\Σ}D(x,q),i=0,1,$

In formula, | C_i| represent bunch C_iSize；C₀Represent the target that the subimage in the Hash bucket hit by query sample is constituted Bunch；C₁Represent by the non-targeted bunch constituted with the subimage in the Hash bucket of query sample arest neighbors；D (x, q) represent sample x with Distance between query sample q, for color characteristic, (x q) uses rectangular histogram to hand over distance to D；For textural characteristics, (x q) adopts D Use Euclidean distance；R_tI () represents the average distance between query sample and each bunch；

Step 5, select the minimum color characteristic of index Validity Index value, textural characteristics as optimum color characteristic and the most respectively Excellent textural characteristics, and determine optimum color characteristic and the binaryzation weight of optimum textural characteristics in accordance with the following methods: index is effectively Property desired value more than 1, then weight is set to 0；Otherwise, weight is set to 1；

Step 6, when two kinds of optimal characteristics exist the feature that weight is 0, only use weight be 1 feature space in LSH Retrieval result is as final remote Sensing Image Retrieval result；When two kinds of optimal characteristics weights are 1, to both optimal characteristics All subimages that LSH retrieval result in space is comprised, by being weighted linearly the similarity of two kinds of optimal characteristics Computing generates comprehensive similarity, and several have the subimage of maximum comprehensive similarity as final distant before therefrom selecting Sense image searching result；Wherein,

Described generate comprehensive similarity by the similarity of two kinds of optimal characteristics being weighted linear operation, specifically according to Lower method:

Given query sample and described subgraph image set is calculated respectively in optimum color feature space and optimum textural feature space In the variance of spacing of each subimage；

Then according to below equation determines optimum color characteristic and the comprehensive similarity weight w of optimum textural characteristics_c、w_t:

$w_c=\frac{{\mathrm{\σ}}_c}{{\mathrm{\σ}}_c+{\mathrm{\σ}}_t},$

$w_t=\frac{{\mathrm{\σ}}_t}{{\mathrm{\σ}}_c+{\mathrm{\σ}}_t},$

Wherein, σ_cAnd σ_tRepresent that in optimum color feature space and optimum textural feature space, given query sample is with described respectively Subimage concentrates the variance of the spacing of each subimage；

Given comprehensive distance D between query sample and arbitrary subimage i is obtained finally according to following formula_i, comprehensive distance is the least, table Bright comprehensive similarity is the highest:

D_i=w_cD_ci+w_tD_ti,

Wherein, D_ciAnd D_tiRepresent given query sample and subgraph in optimum color feature space and optimum textural feature space respectively As the distance between i；

Described therefrom select before several have the subimage of maximum comprehensive similarity as final remote Sensing Image Retrieval result, The most in accordance with the following methods:

Concentrate selected part subimage from described subimage, and divide in optimum color feature space and optimum textural feature space Do not calculate average T of this partial subgraph picture distance each other_cAnd T_t；

Determine similarity threshold T according to the following formula:

T=w_cT_c+w_tT_t,

Wherein, w_cAnd w_tIt is respectively optimum color characteristic and the comprehensive similarity weight of optimum textural characteristics；

In all subimages that last LSH retrieval result from two kinds of optimal characteristics spaces is comprised, choose and given inquiry The comprehensive distance between the sample subimage less than or equal to similarity threshold T, as final remote Sensing Image Retrieval result.

2. the remote sensing image retrieval method combined based on multiple features LSH index as claimed in claim 1, it is characterised in that described The feature space that at least two is different includes the textural characteristics that color feature space that at least two is different is different with at least two Space.

3. the remote sensing image retrieval method combined based on multiple features LSH index as claimed in claim 1, it is characterised in that from institute State subimage concentration selected part subimage to specifically refer to concentrate the subimage randomly selecting 10% from subimage.

4. remote sensing image retrieval method based on multiple features LSH index combination, its feature as described in any one of claims 1 to 3 It is, described piecemeal that remote sensing images are carried out, uses overlap partition strategy, remote sensing images are divided into series of partially overlapping The image block of 64*64 size, between block, overlapping size is 60*60.

5. remote sensing image retrieval method based on multiple features LSH index combination, its feature as described in any one of claims 1 to 3 Being, described step 1, step 2 off-line execution, described step 3 step 6 performs online.

6. remote sensing image retrieval method based on multiple features LSH index combination, its feature as described in any one of claims 1 to 3 It is, when setting up LSH index, hash function quantity M included in Hash table number L, each Hash table and Hash table width The value of degree W is: L=20, M=5, W=4.