CN103902704A - Multi-dimensional inverted index and quick retrieval algorithm for large-scale image visual features - Google Patents

Abstract

The invention discloses a multi-dimensional inverted index and quick retrieval algorithm for large-scale image visual features. The algorithm includes: training a multi-layer codebook required for enhanced residual quantification by visual features of images and building a multi-dimensional inverted index with the trained codebook; according to the trained codebook, quantifying and encoding the image visual features, and according to calculated codes, inserting the codes into corresponding inverted lists in the inverted index; querying the image visual features to query the built multi-dimensional inverted index so as to obtain a query candidate set; optimizing the query candidate set by adaptive hypersphere filtering, and ranking filtered query results so as to finish retrieval of the image visual features. The algorithm has the advantages that as the image features are quantified and encoded, efficiency of quantifying the image features is improved; the multi-dimensional inverted index is built with the generated image codes, the inverted index can be built just by training few codebooks, and speed of building the index structure is increased.

Description

Towards multidimensional inverted index and the quick retrieval of extensive Image Visual Feature

Technical field

The invention belongs to image retrieval technologies field, more specifically, relate to a kind of multidimensional inverted index and quick retrieval towards extensive Image Visual Feature.

Background technology

Current, the development of global network and universal reached unprecedented scale, shares in website multi-medium data every day taking image as representative with surprising speed rapid growth from microblogging, mobile phone, social network sites, news website and multimedia.In the face of the image library of magnanimity, only have and image is organized effectively so that browse, access and retrieve, quick oneself picture interested and that like that also obtain exactly of people's ability.Traditional text needs manually picture material to be marked conventionally, but along with the increase of image library scale, artificial mark is not only time-consuming but also require great effort, and its limitation is more and more obvious.In addition, image text is described and is conventionally derived from the textual description of artificial mark or webpage epigraph, but the textual description of image itself is with strong subjective colo(u)r, thereby CBIR arises at the historic moment.As its name suggests, the picture that it is submitted to according to user, analyzes the visual signature in picture, then searches the picture that comprises similar content, and wherein, how to retrieve the visual signature similar to it according to the visual signature of image is one of key.

Image Visual Feature retrieval mainly comprises index based on tree structure and retrieval, index and retrieval and the inverted index based on vision word and retrieval based on Hash.

In recent years, researchist has carried out a lot of research in index and the retrieval of tree structure both at home and abroad, and obtains good retrieval effectiveness at the lower feature space of dimension, but these methods can face " dimension disaster " in the time processing high dimensional data.So, index based on Hash and retrieval become researchers' focal point, comprising: with accurate European position sensing Hash (Exact Euclidian Locality Sensitive Hashing, E2LSH) for unique point is mapped to low-dimensional theorem in Euclid space by the method for representative, and use Euclidean distance to weigh the similarity between unique point; Unique point be mapped to low-dimensional Hamming space and ensure that unique point similar in theorem in Euclid space has similar binary coding taking spectrum Hash as representative, is used Hamming distance to weigh the similarity between unique point conventionally.These class methods of E2LSH are because needs are stored in internal memory to improve retrieval rate by image feature data, thereby its committed memory space is excessive, has limited manageable database scale.Hash coding calculates similarity by binary coding representation Image Visual Feature and by Hamming distance between coding, can significantly reduce memory space requirements and raising retrieval rate, but, binary-coded length restriction the separating capacity of Hamming distance.The inverted index based on vision word taking word bag model as representative and retrieval are to introduce from text retrieval field, become the study hotspot of field of image search in recent years.First these class methods conventionally quantize and encode a series of method Image Visual Feature, such as: Hamming embedding, accumulated amount, transform coding and residual quantization etc.; Then carry out the metadata such as coding of memory image and improve inquiry velocity by building an inverted index.These class methods had not only kept the low storage demand of Hash coding but also had had advantages of the high separating capacity of Euclidean distance, and then made to inquire about precision and recall precision is all significantly improved.

Although obtained a lot of achievements in research about the retrieval of Image Visual Feature at present, its retrieval performance still has the space of further raising.Such as: how to train code book more accurately for Image Visual Feature being quantized and encoding, to reduce memory space requirements; How to improve the efficiency of characteristic quantification; The inverted index how rapid build comprises fairly large Inverted List; And how further to promote the speed of Image Visual Feature retrieval.

Summary of the invention

The object of the present invention is to provide a kind of multidimensional inverted index and quick retrieval towards extensive Image Visual Feature, be intended to by Image Visual Feature being quantized and encoding, reduce the required space requirement of memory image visual signature; Improve the efficiency of Image Visual Feature insertion inverted index and improve query performance by building multidimensional inverted index; Filter irrelevant Query Result by self-adaptation suprasphere filter algorithm, in the situation that not affecting inquiry accuracy rate, reduce the quantity of sequencing feature, improve the retrieval rate of Image Visual Feature.

Realize the concrete technical scheme that the object of the invention adopts as follows:

Towards multidimensional inverted index and the quick retrieval of extensive Image Visual Feature, by Image Visual Feature being quantized, encode, built inverted index and inquiry, thereby realize the retrieval of image, the method comprises:

The required multilayer code book of visual signature training enhancement mode residual quantization that utilizes image, comprises inceptive code book training and optimizes two stages of code book, and utilizes the code book of training to build multidimensional inverted index;

According to the code book of the enhancement mode residual quantization of having trained, Image Visual Feature is quantized and encoded, be inserted into Inverted List corresponding in inverted index according to the coding calculating simultaneously;

Utilize query image visual signature to inquire about constructed multidimensional inverted index, obtain query candidate collection;

Utilize self-adaptation suprasphere to filter query candidate collection is optimized, to the result ranking after filtering, thereby complete the retrieval of Image Visual Feature.

First the present invention utilizes the required multilayer code book of Image Visual Feature training set training enhancement mode residual quantization and builds inverted index; Then according to the code book of the enhancement mode residual quantization of having trained, utilize the arest neighbors lookup method based on non-linear filtration that Image Visual Feature storehouse is quantized and encoded; Then utilize query image visual signature to inquire about multidimensional inverted index; Finally utilize self-adaptation suprasphere filter method to filter uncorrelated Query Result sequence.Concrete steps are as follows:

(1) code book training builds with multi-dimensional indexing

First, utilize residual quantization method to train L layer code book by k-means method on an Image Visual Feature collection, every layer of code book comprises k cluster center of gravity; Then utilize the method for combined optimization this L layer code book to be optimized to the L layer code book of the new residual quantization that is enhanced.Based on this, before utilizing, the syntagmatic of cluster center of gravity in M layer code book, builds one and comprises at most k ^mthe multidimensional inverted index of individual Inverted List.

(2) Image Visual Feature quantizes and coding

First, utilize the code book of the enhancement mode residual quantization training, successively Image Visual Feature is quantized, obtain L layer coding; Then, according to the front M layer coding of Image Visual Feature, be inserted into Inverted List corresponding in multidimensional inverted index, be saved in ID and L layer coding thereof that content is Image Visual Feature.In addition, in the process that Image Visual Feature is quantized, design the accurate arest neighbors lookup method of a kind of non-linear filtration, on lower dimensional space, utilize the lower limit of Euclidean distance to filter non-neighbour's cluster center of gravity and calculate nearest neighbor classifier center of gravity.

(3) Image Visual Feature inquiry

First, calculate distance between the key word that query image visual signature q and all Inverted Lists are corresponding; Then, Inverted List corresponding to the w of a selected distance minimum key word; Finally feature in this w Inverted List is taken out as query candidate collection.

(4) self-adaptation suprasphere filters and sequence

First, construct a suprasphere taking q as the centre of sphere, its radius calculates to w nearest index key according to q; Then, calculate q and concentrate the distance between all results and will think uncorrelated Query Result apart from being less than the Query Result that suprasphere radius is corresponding with query candidate, only retain the Query Result that is positioned at suprasphere inside; Finally, to the result ranking after filtering, complete the retrieving of Image Visual Feature.

The present invention has designed the search method that enhancement mode residual quantization method, the accurate nearest neighbor classifier center of gravity of non-linear filtration lookup method, multidimensional inverted index build and filter based on self-adaptation suprasphere, improves the performance of Image Visual Feature retrieval.Particularly, the present invention has the following advantages:

(1) improve inquiry accuracy rate and the storage space that reduces Image Visual Feature, the present invention utilizes enhancement mode residual quantization method to quantize Image Visual Feature, make Image Visual Feature obtain more accurate approximate representation, and then improve and inquire about accuracy rate, in addition, quantize the coding that obtains in order to replace Image Visual Feature to be kept in inverted index structure, thereby reduced memory space requirements.

(2) improve Image Visual Feature and quantize and efficiency, the present invention utilizes non-linear filtration to search accurate nearest neighbor classifier center of gravity, has reduced and has calculated the nearest neighbor classifier required time of center of gravity.

(3) improve inverted index and build speed, the present invention only needs a small amount of cluster center of gravity, just can build the inverted index that comprises fairly large Inverted List, in addition, Image Visual Feature only need just can be inserted into corresponding Inverted List according to its front M layer coding, has reduced time overhead.

(4) improve inquiry velocity, the present invention utilizes self-adaptation suprasphere to filter to reduce the quantity of ranking results, reduces the time overhead of sequence, and then in the situation that ensureing inquiry accuracy rate, improves inquiry velocity.

Brief description of the drawings

Fig. 1 is multidimensional inverted index and the quick retrieval process flow diagram of the embodiment of the present invention;

Fig. 2 is the multidimensional inverted index schematic diagram of the embodiment of the present invention;

Fig. 3 is the schematic diagram that the ERVQ of the embodiment of the present invention quantizes and encodes Image Visual Feature;

Fig. 4 is that the self-adaptation suprasphere of the embodiment of the present invention filters schematic diagram.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearer, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein, only in order to explain the present invention, is not intended to limit the present invention.In addition,, in each embodiment of described the present invention, involved technical characterictic just can combine mutually as long as do not form each other conflict.

In the embodiment of the present invention, choose the visual signature of sift feature as image.The embodiment of the present invention is mainly divided into three parts: training module, generates the multilayer code book of enhancement mode residual quantization and utilize these code books to build multidimensional inverted index; Characteristic quantification module: Image Visual Feature is quantized and encoded, be inserted into corresponding Inverted List simultaneously; Enquiry module according to query image visual signature, is searched similar features and is returned in inverted index structure.In the present embodiment, what Image Visual Feature adopted is disclosed data set in the world.

Particularly, as shown in Figure 1, the multidimensional inverted index towards extensive Image Visual Feature and the quick retrieval of the present embodiment specifically comprise following process:

(1) training of the code book of enhancement mode residual quantization and multidimensional inverted index build

Code book training comprises two stages: the combined optimization of inceptive code book training and code book.

(1.1) inceptive code book training

The training process of inceptive code book as shown in algorithm 1, a given Image Visual Feature training set X={x ₁, x ₂... x _i... x _n, first utilize k-means algorithm to obtain k cluster center of gravity to its cluster, obtain the inceptive code book of ground floor; Then, calculation training concentrates proper vector and its residual vector between nearest cluster center of gravity in this layer of code book to obtain E1; And then the training data using E1 as training second layer code book, the clustering method of employing remains k-means algorithm, and then obtains the 2nd layer of code book.If need the code book number of plies L>2 of training, need so again to calculate residual vector cluster.This process is circulated to last one deck inceptive code book trained till, obtain inceptive code book C _l={ c _{l, 1}, c _{l, 2}..., c _l,kl=1,2 ..., L.

In addition, the training process of inceptive code book also will generate some other data for combined optimization, comprising: each layer of code book added up R to the dividing condition of training sample set _l={ R _l,j(j=1,2 ..., k, l=1,2 ..., L), R _l,jrepresent (the 1st layer: X={x of l layer training sample set ₁, x ₂..., x _i... x _n, 2-L layer: E _l-1=ε _{1, l-1}, ε _{2, l-1}..., ε _{n, l-1}) in k cluster center of gravity of this layer of code book in j the sample characteristics point set that cluster center of gravity is minimum distance $R_{l,j}=$ $\{{\mathrm{ID}}_{c_{\mathrm{lj}},1},...,{\mathrm{ID}}_{c_{l,j},n_{l,j}}\},$ Wherein n _{l, j}for falling into C _ljproper vector number; Each layer of quantized result of training feature vector

the overall quantization error MSE of training sample set, its account form is as follows:

$\mathrm{MSE}=E\left[d{(x_j,{\hat{x}}_j)}^2\right]=E{\left|\right|x_j-{\hat{x}}_j\left|\right|}^2=\frac{1}{n}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\underset{i=1}{\overset{d}{\mathrm{\Σ}}}{\left|\right|x_{\mathrm{ji}}-{\hat{x}}_{\mathrm{ji}}\left|\right|}^2$

Wherein,

representation feature vector x _jquantized result; The dimension of d representation feature vector; Proper vector number in n representation feature vector set X.

In algorithm 1, each layer of quantized result of code book C, sample set dividing condition R and each sample characteristics

according to generating from the 1st layer to L layer order.In the data generating procedure of every one deck, first utilize k-means to train and obtain code book C on the input training sample set of this layer _l; Then according to C _l, calculate respectively each training sample proper vector in the quantized result of this layer and add up the spatial division situation of code book to training sample set.After the initialization data of all ERVQ completes, calculate initial overall quantization error MSE.

(1.2) combined optimization of code book

For simplicity, in combined optimization process, the mode of adding subscript (iter) with variable represents the value of variable in the iterth time iteration optimization process, as: code book

sample set dividing condition

the quantized result of sample characteristics collection with overall quantization error MSE ^(iter).What represent when iter=0 variations per hour is initialization data.The initialization data generating corresponding to a upper joint, they are expressed as in this section: $C^{\left(0\right)}=\left\{C_l^{\left(0\right)}\right\},R^{\left(0\right)}=\left\{R_l^{\left(0\right)}\right\},R_l^{\left(0\right)}=\left\{{R_{l,j}}^{\left(0\right)}\right\},{\hat{x}}_j=\{{{\hat{x}}_{i,1}}^{\left(0\right)},{{\hat{x}}_{i,2}}^{\left(0\right)},...,{{\hat{x}}_{i,L}}^{\left(0\right)}\}$ And MSE ⁽⁰⁾.A given L layer, inceptive code book are every layer of ERVQ that code book comprises k cluster center of gravity, the process of its combined optimization is as shown in algorithm 2.

Code book in algorithm 2

combined optimization be according to from the 1st layer to L layer in sequence.In a combined optimization iterative process (3-22 is capable), for the optimization (7-9 is capable) of j cluster center of gravity in l layer code book, the proper vector that falls into cluster center of gravity during first according to last iteration is calculated the residual vector of these vectors and other layer of quantized result; Then these residual vectors are averaged and as j new cluster center of gravity.So circulation is until the k of this layer cluster center of gravity update all is complete.With regard to being equivalent to, l layer is used as to the overall quantization error that last one deck is processed and made the residual vector calculating comprise proper vector like this.After l layer code book upgraded, just training sample proper vector is carried out to re-quantization to upgrade its each layer of quantized value

and

and R _m ^(iter)(10-21 is capable).Here, owing to upgrading the front l-1 layer quantized result that l layer code book can effect characteristics vector, therefore only need regeneration characteristics vector at l layer to the quantized result of L layer and

and R _m ^(iter), reduce the time complexity of training.After an iteration of code book optimization completes, the average overall quantization error of calculation training sample set proper vector (the 23rd row), when

while being less than default threshold value, just think that iteration restrains, finish combined optimization and return to the each layer of up-to-date code book C of ERVQ ^*.Otherwise, continue associating code book optimization until the convergence of average overall quantization error.

(1.3) multidimensional inverted index builds

Utilize the front M layer code book C of ERVQ _l={ c _{l, 1}, c _{l, 2}..., c _l,kl=1,2 ..., M, every layer of code book has k cluster center of gravity, then combines one by one if take out respectively a cluster center of gravity from this M code book, comprises at most k thereby can build ^mthe inverted index structure of individual Inverted List

, as figure) and as shown in 2.The index key of Inverted List is (index ₁..., index _i..., index _m), wherein index _irepresent the cluster center of gravity numbering of i layer code book.The cluster center of gravity that Inverted List is corresponding is defined as the vector sum of M the cluster center of gravity that index key is corresponding ${\mathrm{cw}}_{({\mathrm{<figure-callout\; id="1"\; label="index"\; filenames="HDA0000485286490000012.png"\; state="\{\{state\}\}">index</figure-callout>}}_1,...,{\mathrm{index}}_i,...,{\mathrm{index}}_M)}={\mathrm{<figure-callout\; id="1"\; label="C"\; filenames="HDA0000485286490000012.png"\; state="\{\{state\}\}">C</figure-callout>}}_{1{,\mathrm{<figure-callout\; id="1"\; label="index"\; filenames="HDA0000485286490000012.png"\; state="\{\{state\}\}">index</figure-callout>}}_1}+C_{{2,\mathrm{index}}_2}+...+c_{{M,\mathrm{index}}_M}.$ In the time of M=1, the inverted index structure of structure is exactly an one dimension tabular form; In the time of M=2, the inverted index structure of structure is exactly a bivariate table form, the corresponding Inverted List of each list item; In the time of M=3, the inverted index structure of structure is exactly a three-dimensional cube form, and so on.

(2) visual signature quantizes and insertion inverted index

(2.1) Image Visual Feature quantizes and coding

(2.1.1) step that Image Visual Feature quantizes

Fig. 3 is 2 layers of instance graph that ERVQ quantizes and encodes Image Visual Feature x, the method that it quantizes proper vector employing order.First input feature value x finds the nearest cluster center of gravity of Euclidean distance in the code book of the 1st layer of quantizer, and sets it as this proper vector the quantized result of the 1st layer record x simultaneously and identify ID the quantification of the 1st layer _{x, 1}as the coding of ground floor; Then calculate x with

residual error

and use the same method and calculate it the quantized result of the 2nd layer

with quantification mark ID _{x, 2}; If the number of plies L>2 of ERVQ, so, said process is circulated to L layer successively to be finished, and completes quantification and the coding of Image Visual Feature.

(2.1.2) search nearest cluster center of gravity

Euclidean distance between proper vector x and cluster center of gravity y is defined as with regard to lower limit:

lb(x,y)＝||y|| ²-2d(μ _xμ _y+σ _xσ _y)

Wherein, the dimension that d is proper vector, μ _x, μ _y, σ _x, σ _ybe respectively x and average and standard deviation, concrete account form is as follows:

$\mathrm{\&mu;}=\frac{1}{d}\underset{i=1}{\overset{d}{\mathrm{\&Sigma;}}}{x}_i$

$\mathrm{\&sigma;}=\sqrt{\frac{1}{d}\underset{i=1}{\overset{d}{\mathrm{\&Sigma;}}}{(x_i-\mathrm{\&mu;})}^2}$

Lb is used for filtering non-neighbour's cluster center of gravity, reduces the number of times that calculates Euclidean distance in higher-dimension primitive character space.Accordingly, Euclidean distance between proper vector x and cluster center of gravity y is deducted to the mould of x || x|| ²be defined as the distance between x and y:

$\tilde{d}{(x,y)}^2=d{(x,y)}^2-{\left|\right|x\left|\right|}^2={\left|\right|y\left|\right|}^2-2<x,y>$

A given proper vector x to be quantified and code book C={c _ii=1,2 ..., k(c _irepresent a cluster center of gravity), before algorithm starts, need to carry out some pre-service, comprising: the expectation of calculating each cluster center of gravity in code book C

standard deviation and the expectation μ of Image Visual Feature x to be quantified _xand σ _x.

Algorithm 3 is for utilizing the method for lower limit filtration to find the process of the accurate nearest neighbor classifier center of gravity of proper vector x in k cluster center of gravity.Quantize corresponding to ERVQ, for find the accurate nearest neighbor classifier center of gravity of error originated from input vector in certain one deck code book.Starting before circulation searching, from k cluster center of gravity, select at random a cluster center of gravity as the initial nearest neighbor classifier center of gravity (Seed Points) of x and by distance

as corresponding minor increment min_d (1-2 is capable).In whole algorithmic procedure, use c ^nearestthe current up-to-date accurate nearest neighbor classifier center of gravity (3-15 is capable) of recording feature vector x.As proper vector x and certain cluster center of gravity c _idistance lower limit meet

time, this cluster center of gravity c is just described _ito the Euclidean distance of x

certain satisfied

therefore c _icertainly not accurate nearest neighbor classifier center of gravity and being filtered, without calculating

otherwise, need further calculating

and compare with min_d.

(2.2) insert inverted index

Fig. 2 has shown that image feature vector x quantizes to obtain after coding, according to corresponding mark vector ID through ERVQ by image feature vector x insertion inverted index process equally _x=(ID _{x, 1}..., ID _x,M..., ID _x,L) in before M layer mark (ID _{x, 1}..., ID _x,M), be inserted into (ID in inverted index structure _{x, 1}..., ID _{x, M}) corresponding Inverted List.In Inverted List, preserve is the feature number ID of x and the residue L-M layer mark (ID of ERVQ coding _{x, M+1}..., ID _x,L), its reason is that the front M layer mark of x can obtain by the index key of corresponding Inverted List.

(3) Image Visual Feature retrieval

For query image visual signature q, concrete query steps is as follows:

(3.1) calculate the q index key corresponding with all Inverted Lists

between distance, computing formula is as follows:

$d{(x,\hat{y})}^2={\left|\right|x\left|\right|}^2+{\left|\right|y\left|\right|}^2-2<x,\hat{y}>={\left|\right|x\left|\right|}^2+{\left|\right|\hat{y}\left|\right|}^2-2<x,\stackrel{l=1}{\mathrm{\&Sigma;}}{\hat{y}}_1>={\left|\right|x\left|\right|}^2+{\left|\right|\hat{y}\left|\right|}^2-2\underset{l=1}{\overset{M}{\mathrm{\&Sigma;}}}<x,c_{{l,\mathrm{ID}}_{y,l}}>$

Wherein,

for query feature vector x and keyword corresponding front M strata class center of gravity

inner product;

(3.2) by nearest { d ₁, d ₂..., d _winverted List corresponding to the individual key word of w (w>=1) in Image Visual Feature as Query Result Candidate Set RS _q={ y ₁, y ₂... y _m;

(3.3) for q constructs a suprasphere taking it as the centre of sphere, the computing formula of radius is:

$R_q=\mathrm{\&lambda;}\&times;\frac{1}{w}\underset{i=1}{\overset{w}{\mathrm{\&Sigma;}}}{d}_i$

Wherein, λ is scale-up factor, and for adjusting the size of suprasphere radius, the value in the present embodiment is 1;

(3.4) utilize formula in (3.1) to calculate the distance between all characteristics of image and R in q and RSq _qrelatively, only retain the Query Result meeting with lower inequality, obtain new Query Result RS _qnew=y ' ₁, y ' ₂... y ' _b, Fig. 4 is the schematic diagram that self-adaptation suprasphere filters.

||q-y _i||≤R _q(i=1,2,...,m)

(3.5) according to RS _qnewdistance between middle Image Visual Feature and q sorts to it, and the knn of a layback minimum unique point is as final Query Result.

Those skilled in the art will readily understand; the foregoing is only preferred embodiment of the present invention; not in order to limit the present invention, all any amendments of doing within the spirit and principles in the present invention, be equal to and replace and improvement etc., within all should being included in protection scope of the present invention.

Claims (9)

1. towards multidimensional inverted index and the quick retrieval of extensive Image Visual Feature, by Image Visual Feature being quantized, encode, built inverted index and inquiry, thereby realize the retrieval of image, it is characterized in that, the method comprises:

The required multilayer code book of visual signature training enhancement mode residual quantization that utilizes image, comprises inceptive code book training and optimizes two stages of code book, and utilizes the code book of training to build multidimensional inverted index;

According to the code book of the enhancement mode residual quantization of having trained, Image Visual Feature is quantized and encoded, be inserted into Inverted List corresponding in inverted index according to the coding calculating simultaneously;

Utilize query image visual signature to inquire about constructed multidimensional inverted index, obtain query candidate collection;

Utilize self-adaptation suprasphere to filter query candidate collection is optimized, to the result ranking after filtering, thereby complete the retrieval of Image Visual Feature.

2. multidimensional inverted index and quick retrieval towards extensive Image Visual Feature according to claim 1, is characterized in that, described inceptive code book training process is:

To Image Visual Feature training set, Y carries out cluster, the code book using the cluster center of gravity obtaining as ground floor quantizer;

Obtain approximate vectorial Δ Y by this ground floor quantizer residual error ₁with quantization error E ₁, to this quantization error E ₁carry out cluster, the code book using the cluster center of gravity obtaining as second layer quantizer;

Obtain approximate vectorial Δ Y by this second quantizer residual error ₂with quantization error E ₂;

Loop successively, after multi-layer quantification tolerance, obtain the code book of corresponding quantizer, thereby obtain the required inceptive code book of training enhancement mode residual quantization method; Wherein, described Image Visual Feature training set Y is global characteristics or local feature.

3. multidimensional inverted index and quick retrieval towards extensive Image Visual Feature according to claim 2, is characterized in that, the process of described optimization code book is:

According to the inceptive code book having trained, Image Visual Feature training set Y is quantized to obtain corresponding multi-layer quantification result;

For every one deck code book, utilize the vectorial residual error between Y and the quantized result of its other corresponding all layers, recalculate this layer of code book and upgrade the quantized result of Y at each layer of quantizer;

This optimizing process loops to last one deck code book successively from ground floor code book, optimizes stop condition until meet.

4. according to multidimensional inverted index and quick retrieval towards extensive Image Visual Feature described in claim 2 or 3, it is characterized in that, quantification and the cataloged procedure of Image Visual Feature are as follows:

For Image Visual Feature X, in ground floor code book, find nearest cluster center of gravity as its quantized result and using the ID of this cluster center of gravity as X the coding in ground floor quantizer;

The quantization error producing in ground floor quantizing process is found in second layer code book nearest cluster center of gravity as its quantized result and using the ID of this cluster center of gravity as X the coding in this layer quantizer;

This process loops successively until last one deck quantizer, thereby completes quantification and the coding to X.

5. multidimensional inverted index and quick retrieval towards extensive Image Visual Feature according to claim 4, is characterized in that, in characteristic quantification and cataloged procedure, accurately the lookup method of nearest neighbor classifier center of gravity is as follows:

First, cluster center of gravity in Image Visual Feature vector sum code book be all mapped to a lower dimensional space and in this lower dimensional space, calculate the lower limit of Euclidean distance between proper vector and cluster center of gravity;

Then, filter successively non-neighbour's cluster center of gravity according to the lower limit of Euclidean distance, and then complete accurate nearest neighbor classifier center of gravity and search.

6. multidimensional inverted index and quick retrieval towards extensive Image Visual Feature according to claim 1, is characterized in that, described multidimensional inverted index building process is as follows:

The front M layer that utilizes the multilayer code book of having trained takes out respectively a cluster center of gravity and then combines one by one from this M code book, comprises at most k thereby realize one of structure ^mthe inverted index structure of individual Inverted List, wherein k is the quantity of cluster center of gravity in every layer of code book; The index key that Inverted List is corresponding is the vectorial sum of this M cluster center of gravity.

7. according to multidimensional inverted index and quick retrieval towards extensive Image Visual Feature described in claim 1 to 6 any one, it is characterized in that, the described detailed process of Image Visual Feature being inserted to multidimensional inverted index is as follows:

According to the front M layer coding of the Image Visual Feature calculating, be inserted into Inverted List corresponding in inverted index structure.

8. according to multidimensional inverted index and quick retrieval towards extensive Image Visual Feature described in claim 1 to 7 any one, it is characterized in that, the described query image visual signature that utilizes is inquired about constructed multidimensional inverted index, and concrete steps are as follows:

First, search nearest w (w >=1) the bar Inverted List of multidimensional inverted index middle distance query image visual signature;

Then, the Image Visual Feature in the table of falling row chain is accordingly taken out as Candidate Set.

9. multidimensional inverted index and quick retrieval towards extensive Image Visual Feature according to claim 8, is characterized in that, described utilize self-adaptation suprasphere filter query candidate collection is optimized, process is as follows:

First,, according to the distance of the query image visual signature q index key corresponding with w nearest Inverted List, build a suprasphere taking q as the centre of sphere;

Then, concentrate the Image Visual Feature being positioned at outside suprasphere to filter out query candidate, only retain the Query Result that is positioned at suprasphere inside.

Images