CN103226585A - Self-adaptation Hash rearrangement method for image retrieval - Google Patents

Self-adaptation Hash rearrangement method for image retrieval Download PDF

Info

Publication number
CN103226585A
CN103226585A CN2013101231634A CN201310123163A CN103226585A CN 103226585 A CN103226585 A CN 103226585A CN 2013101231634 A CN2013101231634 A CN 2013101231634A CN 201310123163 A CN201310123163 A CN 201310123163A CN 103226585 A CN103226585 A CN 103226585A
Authority
CN
China
Prior art keywords
image
hash
hash codes
storehouse
class
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.)
Granted
Application number
CN2013101231634A
Other languages
Chinese (zh)
Other versions
CN103226585B (en
Inventor
孔祥维
卢佳音
付海燕
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Dalian University of Technology
Original Assignee
Dalian University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Dalian University of Technology filed Critical Dalian University of Technology
Priority to CN201310123163.4A priority Critical patent/CN103226585B/en
Publication of CN103226585A publication Critical patent/CN103226585A/en
Application granted granted Critical
Publication of CN103226585B publication Critical patent/CN103226585B/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Landscapes

  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention belongs to the technical field of image retrieval, and relates to a self-adaptation Hash rearrangement method for image retrieval, in particular to an image Hash method for carrying out content-based image retrieval. The method adopts the mapping and then sequencing Hash rearrangement method, and comprises the following steps: high-dimensional visual feature vectors of images in a training library are abstracted, a proper Hash method is selected to map the high-dimensional visual feature vectors into Hash codes, and specific class weight vectors are generated for each class of images; the hamming distance between the Hash codes of the retrieved images and the Hash codes in the training library is calculated, and retrieval results are returned in an ascending order; and according to the retrieval results, the self-adaptation weight vectors of the retrieved images are calculated, the hamming distance is weighted by utilizing the structures of the self-adaptation weight vectors of the retrieved images, the returned images are rearranged by utilizing the weighted hamming distance, and more accurate retrieval results are obtained. The self-adaptation Hash rearrangement method calculates specific weights according to the retrieved images, has universality, and remarkably improves the retrieval effect without increase of calculation complexity.

Description

Self-adaptation Hash rearrangement method towards image retrieval
Technical field
The invention belongs to the image retrieval technologies field, relate to and utilize the image hash method to carry out CBIR, specially refer to a kind of self-adaptation Hash rearrangement method towards image retrieval.
Background technology
CBIR is to retrieve as module with the similarity between Image Visual Feature, and its task is to find the image similar in terms of content to retrieving images from image library.Traditional search method European proper vector presentation video of higher-dimension, and utilize the mode of linear sweep that image library is retrieved.Yet when retrieving for the large nuber of images storehouse, because amount of images is huge, corresponding characteristic storage space is huge, and the recall precision of linear sweep mode is very low.And the image hash method is succinct two-value Hash codes with the European Feature Mapping of higher-dimension, has greatly reduced the storage space of feature, and between Hash codes the computing velocity of Hamming distance also obviously faster than the calculating of Euclidean distance; Simultaneously, adopt approximate nearest neighbor method retrieval similar image, also can improve recall precision effectively.
It is as follows to utilize hash method to carry out the key step of image retrieval: the visual signature that at first retrieving images is extracted higher-dimension, select suitable hash method that high dimensional feature is mapped to Hash codes then, the Hamming distance of Hash intersymbol in Hash codes by calculating retrieving images and the image library is returned result for retrieval by ascending order at last.The hash method that the dependence Hamming distance is measured has improved recall precision really, but also has significant disadvantages.Because Hamming distance is an integer, for the large nuber of images storehouse, the distance that has thousands of width of cloth images and retrieving images equates; And Hamming distance only embodied the number of different Hash codes, and ignored the influence of Hash codes position to similarity.And for the image that equates with the retrieving images Hamming distance, the order of returning is extremely important, therefore is necessary the image after the Hash retrieval is reset.
At present, the Hash rearrangement method mainly contains two big classes: mapping back ranking method and ordering back reflection method earlier earlier.Mapping back ranking method is earlier image to be generated Hash codes earlier, and structure weighting Hamming distance is reset returning image.Its feature is to make the Hamming distance serialization, can significantly reduce the amount of images that Hamming distance equates.As Query-Adaptive Ranking[Y.Jiang, J.Wang, and S.Chang.Lost in Binarization:Query-Adaptive Ranking for Similar Image Search with Compact Codes.In proceedings of ICMR, 2011] be the representative of mapping back ranking method earlier.
Ordering back reflection method is to filter out similar image according to the similarity between the European feature of image from image library earlier earlier, then these images is generated Hash codes, resets returning image according to traditional Hamming distance again.As QsRank[X.Zhang, L.Zhang, and H.Shum.QsRank:Query-Sensitive Hash Code Ranking for Efficient ε-neighbor Search.In proceedings of CVPR, 2012] be the typical method of reflection method after the ordering earlier.But because earlier ordering back reflection method utilizes Euclidean distance screening similar image between Image Visual Feature earlier, so time and space complexity are all higher, and recall precision is lower.
Summary of the invention
The technical barrier that the present invention will solve is the defective that overcomes prior art, invent a kind of Hash rearrangement method-self-adaptation Hash rearrangement method that afterwards sorts that shines upon earlier, at first image extracts the visual signature of higher-dimension from the training storehouse, select suitable hash method that high dimensional feature is mapped to Hash codes then, according to the correlativity between the different dimensions of all kinds of Hash codes in the image of training storehouse, obtain the weight vectors of all kinds of images; The Hamming distance of Hash intersymbol in Hash codes by calculating retrieving images and the image library is returned result for retrieval by ascending order again; According to the adaptive weighting of result for retrieval calculating retrieving images, utilize the adaptive weighting vector structure weighting Hamming distance of retrieving images, and utilize the weighting Hamming distance to reset returning image, obtain result for retrieval more accurately.Purpose is to solve the sequencing problem of returning image in the massive image retrieval, has solved the sequencing problem of returning the equal image of image middle distance especially effectively, has improved the accuracy rate of retrieval.
Technical scheme of the present invention is: for the every class image in the training image storehouse, extract the visual feature of image vector, and generate Hash codes, the correlativity between the different dimensions of Hash codes in the class is learnt, for every class image generates specific class weight vectors.For retrieving images, calculate the Hamming distance of image in retrieving images and the image library and return result for retrieval, calculate the adaptive weighting of retrieving images according to result for retrieval.Utilize the weight structure self-adaptation Hamming distance of retrieving images, result for retrieval is reset, obtain result for retrieval more accurately.The specific implementation step comprises:
1, selects retrieving images q, determine image library IM and training storehouse T;
Select retrieving images q, the training storehouse T that determines to include the image library IM of N width of cloth image and comprise M width of cloth image, IM={IM 1, IM 2..., IM N, T={T 1, T 2..., T M,
Wherein: 0＜M≤N;
2, extract visual feature of image, composing images feature database GIM and training characteristics storehouse GT;
Each width of cloth image among image library IM and the training storehouse T utilizes the gist descriptor to extract visual feature of image, and each width of cloth image is represented with the gist proper vector of one 512 dimension; The proper vector composing images feature database GIM of all images among the image library IM, GIM={GIM 1, GIM 2..., GIM N, wherein,
Figure BDA00003029191800031
Expression set of real numbers, the every width of cloth image in each proper vector in the characteristics of image storehouse and the image library are corresponding one by one; The proper vector composing training feature database GT of all images among the training storehouse T, GT={GT 1, GT 2..., GT M, wherein,
Figure BDA00003029191800033
Every width of cloth image in each proper vector in the training characteristics storehouse and the training storehouse is corresponding one by one; The proper vector of retrieving images q is G q,
Figure BDA00003029191800034
3, respectively each proper vector in characteristics of image storehouse and the training characteristics storehouse being generated dimension is the Hash codes of d;
Utilizing existing hash method, is the Hash codes of d to each proper vector generation dimension among characteristics of image storehouse GIM and the training characteristics storehouse GT respectively as LSH, SKLSH or ITQ hash method, is expressed as HI={HI respectively 1, HI 2..., HI NAnd HT={HT 1, HT 2... HT M, wherein HI ∈ 0,1} N * dBe the matrix of N * d dimension, each element of matrix is 0 or 1; HT ∈ 0,1} M * dBe the matrix of M * d dimension, each element of matrix is 0 or 1; If comprise k class image altogether among the training storehouse T, wherein k is a positive integer, then trains the Hash codes HT of storehouse T also can be expressed as by its classification
Figure BDA00003029191800035
Wherein
Figure BDA00003029191800036
The Hash codes set of i class all images, i ∈ [1, k] here among the expression training storehouse T; I class with the training storehouse is an example, and its Hash codes set can be expressed as
Figure BDA00003029191800037
For
Figure BDA00003029191800038
Matrix, each matrix element is 0 or 1 Hash codes, wherein The picture number that comprised of i class image for training storehouse T;
4, the image training of training storehouse is obtained the class weight vectors;
By relatively training i class Hash codes among the T of storehouse
Figure BDA00003029191800041
In each column vector, add up every column mean and be 0 and 1 number, be designated as respectively
Figure BDA00003029191800042
With
Figure BDA00003029191800043
I class Hash codes among the expression training storehouse T
Figure BDA00003029191800044
In 0 and 1 number on the r dimension Hash codes, wherein r ∈ [1, d]; I class Hash codes among the T of calculation training storehouse
Figure BDA00003029191800045
Corresponding class weight vectors
Figure BDA00003029191800046
Wherein
Figure BDA00003029191800047
Be the vector of d dimension, each element in the vector is less than 1 decimal greater than 0; Order max _ num c i , r = max { num 0 c i , r , num 1 c i , r } , Represent i class Hash codes
Figure BDA00003029191800049
In r list 0 or 1 maximum number, then have: max _ num c i = { max _ num c i , 1 , max _ num c i , 2 , . . . , max _ num c i , d } , Expression respectively list 0 or 1 maximum number; Order
Figure BDA000030291918000412
Figure BDA000030291918000413
Be the vector of d dimension, each element in the vector is the decimal between 0.5 to 1, the otherness among the expression training storehouse T between i class Hash codes; According to the otherness between i class Hash codes
Figure BDA000030291918000414
Calculate the corresponding weight vectors of i class Hash codes r dimension: when dp c i , r ∈ [ th j , th j - 1 ) , J=1 ... during L, ω c i , r = ω s j ω _ norm , Wherein,
Figure BDA000030291918000417
Be vector
Figure BDA000030291918000418
R element; Th={th 1... th LBe according to image library preset threshold value vector, each element is the decimal between 0.5 to 1;
Figure BDA000030291918000419
Be the predetermined weights vector, each element is the decimal between 0 to 1; L is a positive integer, is the number of default weight;
Figure BDA000030291918000420
It is the weight vectors of i class Hash codes
Figure BDA000030291918000421
R element;
Figure BDA000030291918000422
Be normalized parameter, realize the normalization of weight vectors; Therefore, i class Hash codes among the training storehouse T
Figure BDA000030291918000423
Corresponding weight vectors is
Figure BDA000030291918000424
I ∈ [1, k] wherein;
5, calculate the adaptive weighting vector of retrieving images q;
Calculate the Hash codes h of retrieving images q earlier qHash codes h with image e in the image library eBetween Hamming distance
Figure BDA00003029191800051
Wherein
Figure BDA00003029191800052
Be the XOR between the scale-of-two Hash codes, dist HammIt is the integer between 0 to d.According to dist HammAscending corresponding image in the image library is sorted, take out and come the most preceding TN width of cloth image, TN is a positive integer here, and represents that with S set C this TN width of cloth image corresponding class gathers,
Figure BDA00003029191800053
Represent to belong in this TN width of cloth image the number of i class image, then the adaptive weighting computing formula of retrieving images q is
Figure BDA00003029191800054
Wherein,
Figure BDA00003029191800055
Be i class Hash codes among the training storehouse T
Figure BDA00003029191800056
Corresponding weight vectors;
6, structure self-adaptation Hamming distance is reset result for retrieval;
The Hash codes h of retrieving images q qHash codes h with image e in the image library eBetween the self-adaptation Hamming distance be defined as:
Figure BDA00003029191800057
Wherein represent the Hadamard product between vector, promptly two corresponding elements of vector multiply each other; According to dist QARAscending the image that returns is reset, obtained result for retrieval more accurately.
But extraction list of references [Aude Oliva about the gist proper vector, Antonio Torralba, Modeling the shape of the scene:a holistic representation of the spatial envelope, International Journal of Computer Vision, Vol.42 (3): 145-175,2001].
Effect of the present invention and benefit are: the present invention proposes a kind of self-adaptation Hash rearrangement method towards image retrieval, utilizes the adaptive weighting vector structure weighting Hamming distance of retrieving images, and utilizes the weighting Hamming distance to reset returning image.This self-adaptation rearrangement method calculates specific weight according to different retrieving images, has generality, and obviously improved retrieval effectiveness when not increasing computation complexity.
Description of drawings
Fig. 1 is the schematic flow sheet towards the self-adaptation Hash rearrangement method of image retrieval that the present invention proposes.
Fig. 2 is when the Hash dimension is 128, utilize the present invention that the result for retrieval of LSH hash method is reset before and reset after return the comparison diagram as a result of preceding 10 width of cloth images.
Fig. 3 is when the Hash dimension is 128, utilize the present invention that the result for retrieval of ITQ hash method is reset before and reset after return the comparison diagram as a result of preceding 10 width of cloth images.
Fig. 4 (a) is before the present invention resets the result for retrieval of LSH hash method and the comparison diagram as a result after resetting; Fig. 4 (b) is before the present invention resets the result for retrieval of SKLSH hash method and the comparison diagram as a result after resetting; Fig. 4 (c) is before the present invention resets the result for retrieval of ITQ hash method and the comparison diagram as a result after resetting.Wherein: horizontal ordinate is represented the Hash dimension, and ordinate represents to return the average retrieval rate of preceding 50 width of cloth images.
Fig. 5 (a) is before the present invention resets the result for retrieval of LSH hash method and the comparison diagram as a result after resetting; Fig. 5 (b) is before the present invention resets the result for retrieval of SKLSH hash method and the comparison diagram as a result after resetting; Fig. 5 (c) is before the present invention resets the result for retrieval of ITQ hash method and the comparison diagram as a result after resetting.Wherein: horizontal ordinate is represented the picture number returned, and ordinate is represented average retrieval rate.
Embodiment
Be described in detail the specific embodiment of the present invention below in conjunction with technical scheme and accompanying drawing.At present, the back of the mapping earlier ranking method in the Hash rearrangement method is earlier image to be generated Hash codes, and structure weighting Hamming distance is reset returning image.Its feature is to make the Hamming distance serialization, can significantly reduce the amount of images that Hamming distance equates.But also there is significant disadvantages.For example, the Hash codes h of known three width of cloth images 1=010, h 2=110, h 3=011, h wherein 1Be the Hash codes of retrieving images, though h 1And h 2Hamming distance and h 1And h 3Hamming distance equate, if but hash function makes first Hash codes more important than the 3rd when mapping, then Hash codes is h 2Image should be h in Hash codes 3Image before be returned.Therefore, for the image that equates with the retrieving images Hamming distance, the order of returning is extremely important.The present invention adopts the elder generation's mapping back ranking method in the Hash rearrangement method, and proposes self-adaptation Hash rearrangement method, solves the sequencing problem of returning the equal image of image middle distance effectively.
Example 1, the coloured image that comprises 10000 100 * 100 pixels in the image library, totally 100 classes, every class 100 width of cloth, deriving from " Product Image Categorization DataSet " commodity image library of people's foundation such as the Xie Xing of Microsoft Research, Asia, is the image library of generally acknowledging in the world.
Step 1. is taken out 1000 width of cloth images at random respectively as retrieving images q from image library, all the other 9000 width of cloth images are as the training storehouse.
Step 2. is converted into gray level image with all coloured images in retrieving images and the training storehouse, extracts the gist visual signature of 512 dimensions.Characteristics of image storehouse and training characteristics storehouse are respectively GIM={GIM 1, GIM 2..., GIM 10000And GT={GT 1, GT 2..., GT 9000, wherein
Figure BDA00003029191800071
Figure BDA00003029191800072
Wherein, gist Feature Extraction process can adopt disclosed matlab code.
Step 3. is utilized disclosed matlab code, selects LSH, SKLSH and these three kinds of hash methods commonly used of ITQ to training characteristics storehouse GT={GT 1, GT 2..., GT 9000To generate dimension be the Hash codes HT={HT of d 1, HT 2... HT 9000.Because training comprises 100 class images in the storehouse, so category also can be expressed as the Hash codes in training storehouse H c = { H c 1 , H c 2 , . . . , H c 100 } .
Step 4. is worked as i=1, and 2 ...,, add up i class Hash codes respectively at 100 o'clock
Figure BDA00003029191800074
R row in 0 and 1 number
Figure BDA00003029191800075
With
Figure BDA00003029191800076
R=1,2 ..., d.In this retrieval, Hash codes dimension d=8,16,32,64,128,256.
Step 5. is calculated i class Hash codes
Figure BDA00003029191800077
Corresponding weight vectors I ∈ [1, k]:
1) gets
Figure BDA00003029191800079
With
Figure BDA000030291918000710
Maximal value, that is:
max _ num c i , r = max { num 0 c i , r , num 1 c i , r } - - - ( 1 )
2) calculate
Figure BDA000030291918000712
respectively list 0 or 1 maximum number:
max _ num c i = { max _ num c i , 1 , max _ num c i , 2 , . . . , max _ num c i , d } - - - ( 2 )
3) otherness between calculating i class Hash codes:
dp c i = max _ num c i m c i - - - ( 3 )
4) setting parameter L=6, th={th 1... th 6}={ 1,0.9,0.8,0.7,0.6,0.5}, ω s = { ω s 1 , . . . , ω s 6 } = { 1,0.8,0.6,0.4,0.3,0.1 } , Normalized parameter then
Figure BDA00003029191800081
So the weight vectors of i class Hash codes r dimension, r ∈ [1, d] is expressed as:
ω c i , r = ω s j ω _ norm = 1 3.2 , dp c i , r = 1 0.8 3.2 , dp c i , r ∈ [ 0.9,1 ) 0.6 3.2 , dp c i , r ∈ [ 0.8,0.9 ) 0.4 3.2 , dp c i , r ∈ [ 0.7,0.8 ) 0.3 3.2 , dp c i , r ∈ [ 0.6,0.7 ) 0.1 3.2 , dp c i , r ∈ [ 0.5,0.6 ) - - - ( 4 )
So i class Hash codes Corresponding weight vectors is
Figure BDA00003029191800084
I ∈ [1, k].
Step 6. is calculated the Hash codes h of retrieving images q qHash codes h with image t in the image library tBetween Hamming distance And press dist HammAscending corresponding image in the image library is sorted, make TN=10, promptly take out and come 10 the most preceding width of cloth images, and therefrom select and comprise three maximum class c of amount of images 1, c 2, c 3, corresponding S set C, the picture number of this three classes correspondence is designated as
Figure BDA00003029191800088
Then the adaptive weighting of retrieving images q is ω q = n c 1 ω c 1 + n c 2 ω c 2 + n c 3 ω c 3 n c 1 + n c 2 + n c 3 .
The ω that step 7. utilizes step 6 to obtain q, can construct the self-adaptation Hamming distance And the image that returns is reset according to its size.
Fig. 2 is when selecting the LSH hash method to generate 128 dimension Hash codes, when returning preceding 10 width of cloth images, and the result's contrast before the present invention resets and after resetting; Fig. 3 is when selecting the ITQ hash method to generate 128 dimension Hash codes, when returning preceding 10 width of cloth images, and the result's contrast before the present invention resets and after resetting.As can be seen, the result before the retrieval effectiveness of the rearrangement method that the present invention proposes obviously is better than resetting has improved retrieval rate from Fig. 2 and Fig. 3.Fig. 4 works as the Hash dimension not simultaneously, returns the average retrieval rate figure of preceding 50 width of cloth images before and after the present invention resets the result for retrieval of three kinds of hash methods.As can be seen from the figure, along with the increase of Hash dimension, retrieval rate is more and more higher, and it is also increasing that the present invention resets the accuracy rate that improves the back.Fig. 5 is the average retrieval rate figure before and after return picture number not simultaneously, the present invention resets the result for retrieval of three kinds of hash methods.Experimental result shows that to return picture number many more, and retrieval rate is low more, and the accuracy rate that the present invention improves is also big more.
The explanation of above example, the present invention can fine solution equate the sequencing problem of image with the retrieving images Hamming distance, improved retrieval rate, can present the better retrieval effect for the user.

Claims (1)

1. self-adaptation Hash rearrangement method towards image retrieval, it is characterized in that, adopt and shine upon earlier the Hash rearrangement method that afterwards sorts, at first extract the higher-dimension visual feature vector of image in the training storehouse, and select suitable hash method that the higher-dimension visual signature is mapped to Hash codes, according to the correlativity between the different dimensions of all kinds of Hash codes in the image of training storehouse, for every class image generates specific class weight vectors; By the Hash codes of calculating retrieving images and the Hamming distance of training Hash intersymbol in the storehouse, return result for retrieval again by ascending order; According to the adaptive weighting vector of result for retrieval calculating retrieving images, utilize the adaptive weighting vector structure weighting Hamming distance of retrieving images, and utilize the weighting Hamming distance to reset returning image, obtain result for retrieval more accurately; Concrete steps are as follows:
1), select retrieving images q, determine image library IM and training storehouse T;
Select retrieving images q, the training storehouse T that determines to include the image library IM of N width of cloth image and comprise M width of cloth image, i.e. IM={IM 1, IM 2..., IM N, T={T 1, T 2..., T M,
Wherein: 0＜M＜N;
2), extract visual feature of image, composing images feature database GIM and training characteristics storehouse GT;
Each width of cloth image among image library IM and the training storehouse T utilizes the gist descriptor to extract visual feature of image, and each width of cloth image is represented with the gist proper vector of one 512 dimension; The proper vector composing images feature database GIM of all images among the image library IM, GIM={GIM 1, GIM 2..., GIM N, wherein,
Figure FDA00003029191700012
Expression set of real numbers, the every width of cloth image in each proper vector in the characteristics of image storehouse and the image library are corresponding one by one; The proper vector composing training feature database GT of all images among the training storehouse T, GT={GT 1, GT 2..., GT M, wherein,
Figure FDA00003029191700013
Every width of cloth image in each proper vector in the training characteristics storehouse and the training storehouse is corresponding one by one; The proper vector of retrieving images q is G q,
Figure FDA00003029191700014
3), respectively each proper vector in characteristics of image storehouse and the training characteristics storehouse being generated dimension is the Hash codes of d;
Utilizing existing hash method, is the Hash codes of d to each proper vector generation dimension among characteristics of image storehouse GIM and the training characteristics storehouse GT respectively as LSH, SKLSH or ITQ hash method, is expressed as HI={HI respectively 1, HI 2..., HI NAnd HT={HT 1, HT 2... HT M, wherein HI ∈ 0,1} N * dBe the matrix of N * d dimension, each element of matrix is 0 or 1; HT ∈ 0,1} M * dBe the matrix of M * d dimension, each element of matrix is 0 or 1; If comprise k class image altogether among the training storehouse T, wherein k is a positive integer, then trains the Hash codes HT of storehouse T also can be expressed as by its classification
Figure FDA00003029191700021
Wherein The Hash codes set of i class all images, i ∈ [1, k] here among the expression training storehouse T; I class with the training storehouse is an example, and its Hash codes set can be expressed as For
Figure FDA00003029191700024
Matrix, each matrix element is 0 or 1 Hash codes, wherein
Figure FDA00003029191700025
The picture number that comprised of i class image for training storehouse T;
4), the image training of training storehouse is obtained class weight vectors ω c
By relatively training i class Hash codes among the T of storehouse
Figure FDA00003029191700026
In each column vector, add up every column mean and be 0 and 1 number, be designated as respectively With
Figure FDA00003029191700028
I class Hash codes among the expression training storehouse T
Figure FDA00003029191700029
In 0 and 1 number on the r dimension Hash codes, wherein r ∈ [1, d]; I class Hash codes among the T of calculation training storehouse Corresponding class weight vectors
Figure FDA000030291917000211
Wherein
Figure FDA000030291917000212
Be the vector of d dimension, each element in the vector is less than 1 decimal greater than 0; Order max _ num c i , r = max { num 0 c i , r , num 1 c i , r } , Represent i class Hash codes
Figure FDA000030291917000214
In r list 0 or 1 maximum number, then have:
max _ num c i = { max _ num c i , 1 , max _ num c i , 2 , . . . , max _ num c i , d } , Expression
Figure FDA000030291917000216
respectively list 0 or 1 maximum number; Order
Figure FDA000030291917000217
Figure FDA000030291917000218
Be the vector of d dimension, each element in the vector is the decimal between 0.5 to 1, the otherness among the expression training storehouse T between i class Hash codes; According to the otherness between i class Hash codes
Figure FDA000030291917000219
Calculate the corresponding weight vectors of i class Hash codes r dimension: when dp c i , r ∈ [ th j , th j - 1 ) , J=1 ... during L, ω c i , r = ω s j ω _ norm , Wherein,
Figure FDA000030291917000222
Be vector
Figure FDA000030291917000223
R element; Th={th 1... th LBe according to image library preset threshold value vector, each element is the decimal between 0.5 to 1;
Figure FDA00003029191700031
Be the predetermined weights vector, each element is the decimal between 0 to 1; L is a positive integer, is the number of default weight;
Figure FDA00003029191700032
It is the weight vectors of i class Hash codes
Figure FDA00003029191700033
R element;
Figure FDA00003029191700034
Be normalized parameter, realize the normalization of weight vectors; Therefore, i class Hash codes among the training storehouse T
Figure FDA00003029191700035
Corresponding weight vectors is
Figure FDA00003029191700036
I ∈ [1, k] wherein;
5), calculate the adaptive weighting vector of retrieving images q;
Calculate the Hash codes h of retrieving images q earlier qHash codes h with image e in the image library eBetween Hamming distance Wherein
Figure FDA00003029191700038
Be the XOR between the scale-of-two Hash codes, dist HammBe the integer between 0 to d, according to dist HammAscending corresponding image in the image library is sorted, take out and come the most preceding TN width of cloth image, TN is a positive integer here, and represents that with S set C this TN width of cloth image corresponding class gathers,
Figure FDA00003029191700039
Represent to belong in this TN width of cloth image the number of i class image, then the adaptive weighting computing formula of retrieving images q is Wherein,
Figure FDA000030291917000311
Be i class Hash codes among the training storehouse T
Figure FDA000030291917000312
Corresponding weight vectors;
6), structure self-adaptation Hamming distance, result for retrieval is reset;
The Hash codes h of retrieving images q qHash codes h with image e in the image library eBetween the self-adaptation Hamming distance be defined as: Wherein represent the Hadamard product between vector, promptly two corresponding elements of vector multiply each other; According to dist QARAscending the image that returns is reset, obtained result for retrieval more accurately.
CN201310123163.4A 2013-04-10 2013-04-10 Towards the self-adaptation Hash rearrangement method of image retrieval Expired - Fee Related CN103226585B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201310123163.4A CN103226585B (en) 2013-04-10 2013-04-10 Towards the self-adaptation Hash rearrangement method of image retrieval

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201310123163.4A CN103226585B (en) 2013-04-10 2013-04-10 Towards the self-adaptation Hash rearrangement method of image retrieval

Publications (2)

Publication Number Publication Date
CN103226585A true CN103226585A (en) 2013-07-31
CN103226585B CN103226585B (en) 2015-10-28

Family

ID=48837031

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201310123163.4A Expired - Fee Related CN103226585B (en) 2013-04-10 2013-04-10 Towards the self-adaptation Hash rearrangement method of image retrieval

Country Status (1)

Country Link
CN (1) CN103226585B (en)

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104951559A (en) * 2014-12-30 2015-09-30 大连理工大学 Binary code rearrangement method based on bit weight
CN105095208A (en) * 2014-04-18 2015-11-25 华为技术有限公司 Method and device for matrix data index
CN106708823A (en) * 2015-07-20 2017-05-24 阿里巴巴集团控股有限公司 Search processing method, apparatus and system
CN106777388A (en) * 2017-02-20 2017-05-31 华南理工大学 A kind of multilist hashing image search method of dual compensation
CN104951791B (en) * 2014-03-26 2018-10-09 华为技术有限公司 data classification method and device
CN109558890A (en) * 2018-09-30 2019-04-02 天津大学 Zero sample image classification method of confrontation network is recycled based on adaptive weighting Hash
CN109634953A (en) * 2018-11-07 2019-04-16 宁波大学 A kind of weight quantization Hash search method towards higher-dimension large data sets
CN109710607A (en) * 2018-11-07 2019-05-03 宁波大学 A kind of hash query method solved based on weight towards higher-dimension big data
CN110059206A (en) * 2019-03-29 2019-07-26 银江股份有限公司 A kind of extensive hashing image search method based on depth representative learning
WO2020098296A1 (en) * 2018-11-15 2020-05-22 ***股份有限公司 Image retrieval method and device
CN112712090A (en) * 2019-10-24 2021-04-27 北京易真学思教育科技有限公司 Image processing method, device, equipment and storage medium
CN113407756A (en) * 2021-05-28 2021-09-17 山西云时代智慧城市技术发展有限公司 Lung nodule CT image reordering method based on self-adaptive weight

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109992716B (en) * 2019-03-29 2023-01-17 电子科技大学 Indonesia similar news recommendation method based on ITQ algorithm

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7672976B2 (en) * 2006-05-03 2010-03-02 Ut-Battelle, Llc Method for the reduction of image content redundancy in large image databases
CN101710334A (en) * 2009-12-04 2010-05-19 大连理工大学 Large-scale image library retrieving method based on image Hash
CN102508910A (en) * 2011-11-11 2012-06-20 大连理工大学 Image retrieval method based on minimum projection errors of multiple hash tables

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7672976B2 (en) * 2006-05-03 2010-03-02 Ut-Battelle, Llc Method for the reduction of image content redundancy in large image databases
CN101710334A (en) * 2009-12-04 2010-05-19 大连理工大学 Large-scale image library retrieving method based on image Hash
CN102508910A (en) * 2011-11-11 2012-06-20 大连理工大学 Image retrieval method based on minimum projection errors of multiple hash tables

Cited By (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104951791B (en) * 2014-03-26 2018-10-09 华为技术有限公司 data classification method and device
CN105095208A (en) * 2014-04-18 2015-11-25 华为技术有限公司 Method and device for matrix data index
CN104951559A (en) * 2014-12-30 2015-09-30 大连理工大学 Binary code rearrangement method based on bit weight
CN106708823A (en) * 2015-07-20 2017-05-24 阿里巴巴集团控股有限公司 Search processing method, apparatus and system
CN106777388B (en) * 2017-02-20 2020-11-24 华南理工大学 Double-compensation multi-table Hash image retrieval method
CN106777388A (en) * 2017-02-20 2017-05-31 华南理工大学 A kind of multilist hashing image search method of dual compensation
CN109558890A (en) * 2018-09-30 2019-04-02 天津大学 Zero sample image classification method of confrontation network is recycled based on adaptive weighting Hash
CN109558890B (en) * 2018-09-30 2023-03-31 天津大学 Zero sample image classification method based on adaptive weight Hash loop countermeasure network
CN109634953B (en) * 2018-11-07 2021-08-17 宁波大学 Weighted quantization Hash retrieval method for high-dimensional large data set
CN109710607A (en) * 2018-11-07 2019-05-03 宁波大学 A kind of hash query method solved based on weight towards higher-dimension big data
CN109710607B (en) * 2018-11-07 2021-09-17 宁波大学 Hash query method for high-dimensional big data based on weight solving
CN109634953A (en) * 2018-11-07 2019-04-16 宁波大学 A kind of weight quantization Hash search method towards higher-dimension large data sets
WO2020098296A1 (en) * 2018-11-15 2020-05-22 ***股份有限公司 Image retrieval method and device
CN110059206A (en) * 2019-03-29 2019-07-26 银江股份有限公司 A kind of extensive hashing image search method based on depth representative learning
CN112712090A (en) * 2019-10-24 2021-04-27 北京易真学思教育科技有限公司 Image processing method, device, equipment and storage medium
CN113407756A (en) * 2021-05-28 2021-09-17 山西云时代智慧城市技术发展有限公司 Lung nodule CT image reordering method based on self-adaptive weight
CN113407756B (en) * 2021-05-28 2022-10-11 山西云时代智慧城市技术发展有限公司 Lung nodule CT image reordering method based on self-adaptive weight

Also Published As

Publication number Publication date
CN103226585B (en) 2015-10-28

Similar Documents

Publication Publication Date Title
CN103226585A (en) Self-adaptation Hash rearrangement method for image retrieval
CN101710334B (en) Large-scale image library retrieving method based on image Hash
US20220222918A1 (en) Image retrieval method and apparatus, storage medium, and device
Zhang et al. Image classification using spatial pyramid robust sparse coding
CN102890700A (en) Method for retrieving similar video clips based on sports competition videos
US20220414144A1 (en) Multi-task deep hash learning-based retrieval method for massive logistics product images
Chen et al. Measuring the effectiveness of various features for thematic information extraction from very high resolution remote sensing imagery
CN103020321B (en) Neighbor search method and system
Liu et al. A novel feature fusion approach for VHR remote sensing image classification
Ye et al. Query-adaptive remote sensing image retrieval based on image rank similarity and image-to-query class similarity
Chen et al. Class-specific feature selection with local geometric structure and discriminative information based on sparse similar samples
CN103235955A (en) Extraction method of visual word in image retrieval
Li et al. A cross-layer nonlocal network for remote sensing scene classification
Li et al. Spatial and temporal information fusion for human action recognition via Center Boundary Balancing Multimodal Classifier
Fierro-Radilla et al. An effective visual descriptor based on color and shape features for image retrieval
Zhu et al. Multi-feature probability topic scene classifier for high spatial resolution remote sensing imagery
Wang et al. Classification of remotely sensed images using an ensemble of improved convolutional network
Wu et al. Ground-based vision cloud image classification based on extreme learning machine
Ahmad et al. A fusion of labeled-grid shape descriptors with weighted ranking algorithm for shapes recognition
Zheng et al. An end-to-end image retrieval system Based on gravitational field deep learning
Chathurani et al. Content-based image (object) retrieval with rotational invariant bag-of-visual words representation
CN104615994B (en) A kind of monitoring image real-time processing method
Fan et al. Robust visual tracking via bag of superpixels
Saini et al. Effects of distance metrics on color recognition using color histogram and KNN algorithm
CN108427967B (en) Real-time image clustering 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
C14 Grant of patent or utility model
GR01 Patent grant
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20151028

Termination date: 20210410

CF01 Termination of patent right due to non-payment of annual fee