CN107423768A - The image Hash sequence generating method combined based on SURF and PCA - Google Patents

The image Hash sequence generating method combined based on SURF and PCA Download PDF

Info

Publication number
CN107423768A
CN107423768A CN201710652338.9A CN201710652338A CN107423768A CN 107423768 A CN107423768 A CN 107423768A CN 201710652338 A CN201710652338 A CN 201710652338A CN 107423768 A CN107423768 A CN 107423768A
Authority
CN
China
Prior art keywords
mrow
image
characteristic point
point
msub
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201710652338.9A
Other languages
Chinese (zh)
Inventor
陈颖
乔君
高乐莲
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.)
Shanghai Institute of Technology
Original Assignee
Shanghai Institute 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 Shanghai Institute of Technology filed Critical Shanghai Institute of Technology
Priority to CN201710652338.9A priority Critical patent/CN107423768A/en
Publication of CN107423768A publication Critical patent/CN107423768A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2136—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on sparsity criteria, e.g. with an overcomplete basis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides a kind of image Hash sequence generating method combined based on SURF and PCA, including：The characteristic point of input picture is detected in metric space；Determine the principal direction of characteristic point；It is divided into n × n block region centered on characteristic point；Extract multiple sample points in each block region both horizontally and vertically on small echo response sum, and to small echo response absolute value sum, obtain the feature description vectors of input picture；The feature description vectors of input picture are subjected to dimension-reduction treatment；Hash sequence is generated according to dimensionality reduction feature description vectors.Method of the invention by extending comparison point number in three dimensions, counted out to reduce low quality feature, improve image recognition rate；Dimensionality reduction is carried out to characteristic point using PCA algorithms, the data after dimensionality reduction are capable of the principal character of representative image, so as to effectively lift the discrimination of image, have good effect in terms of prominent otherness, can clearly give expression to the difference of two images.

Description

The image Hash sequence generating method combined based on SURF and PCA
Technical field
The present invention relates to safety of image technical field, in particular it relates to be based on rapid robust feature (Speeded-Up Robust Features, SURF) and principal component analysis (Principal Component Analysis, PCA) combine image Hash sequence generating method.
Background technology
The nearest more than ten years, with the progress of new and high technology and the fast development of Digital Media, various multimedia databases Using more and more frequent.Meanwhile the growth of multi-medium data causes the storage of data and management to be faced with a series of severe peaces Full problem.In face of these safety problems, image perception Hash is because of its wide applicability, security, robustness by increasingly Known to more people.Original image by normal digital processing (such as geometric transformation, watermark processing, form conversion, at brightness and color Reason etc.) after obtain image to be matched, how to judge whether image and original image to be matched are same image, have become image peace One important research content in full field.
Image Hash has sane characteristic (robustness) to vision, so it is otherwise known as " perception Hash ".Image Hash According to the characteristics of human visual system, it is expressed as to one section short and small of character or Serial No..In simple terms, the present invention is ground The image hash method studied carefully actually be equivalent to study a kind of Image Compression.When in the vision of original image and image to be matched Hold similar, then the Hash sequence of this two images is identical or approximate, and this point embodies the robustness of image Hash.If two width Image vision content difference is bigger, and the Hash sequence of this two images is almost entirely different, and this embodies image Hash again Uniqueness.
Robustness and uniqueness are the most basic performance indications of image Hash, and in addition, image Hash also has oneself Particular characteristic requirement, referred to as security, it is mainly reflected in distorted image detection and safety of image application.
In terms of the progress of research, foreign countries' starting is more early than domestic.The early 1990s, Columbia University Chang etc. is studied Personnel just propose some practical image hash methods.For example image histogram is subjected to hash method as point of penetration and ground Study carefully.Then distorted and normal JPEG (Joint Photographic Experts on malice can be distinguished Group, JPEG) the image hash method of format compression also gradually grows up.Different, the researcher according to Feature Extraction Technology Discrete cosine transform (Discrete Cosine Transform, DCT), wavelet transform (Discrete Wavelet Transformation, DWT) and the introducing of singular value decomposition (Singular Value Decomposition, SVD) scheduling algorithm Image Hash is studied.Recent years, B.Macq professors propose a kind of on Radon changes in European signal transacting meeting The image hash method changed, then such as it is based on also in constantly improve and improvement on the hash method based on being converted by Radon Rash hash method etc..Fourier-Mellin Transform (Fourier-Mellin transform, FMT) structural map as Hash for Geometric transformation also has sane characteristic, and is widely used in medical image, there is very big practicality.Monga et al. is again by square The thinking of battle array dimensionality reduction introduces image Hash, it is proposed that with Non-negative Matrix Factorization (Non-negative Matrix Factorization, NMF) method constructs Hash, also achieve good effect in resistance image geometry attack.
In recent years, the hash method based on characteristics of image turned into study hotspot, was such as based on Scale invariant features transform The hash method of (Scale invariant feature Transform, SIFT), but the characteristic point ratio of SIFT algorithms extraction It is more, and match inaccurate, operation time ratio acceleration robust features (Speeded-Up Robust Features, SURF) calculation Method is slow nearly three times.
Be currently that image Hash develops the stage rapidly, although domestic and foreign scholars have devised it is all kinds of towards practical application Hash algorithm, but the characteristic point of existing image hash algorithm extraction is more, computing is complicated.
The content of the invention
For in the prior art the defects of, it is an object of the invention to provide a kind of image Kazakhstan combined based on SURF and PCA Uncommon sequence generating method.
According to the image Hash sequence generating method provided by the invention combined based on SURF and PCA, including：
The characteristic point of input picture is detected in metric space；The input picture includes：Image to be matched And original image；
Determine the principal direction of the characteristic point；
Centered on the characteristic point, zoning is determined, the zoning is divided into n × n block region；
Extract multiple sample points in each block region both horizontally and vertically on small echo response sum, and to described The absolute value summation of small echo response, obtains the feature description vectors of the input picture；
The feature description vectors of the input picture are carried out by dimension-reduction treatment by principal component analysis PCA algorithms, dropped Dimensional feature description vectors；
Hash sequence is generated according to the dimensionality reduction feature description vectors.
Alternatively, the feature description vectors of the input picture are being carried out at dimensionality reduction by principal component analysis PCA algorithms Before reason, in addition to：
Feature similarity measure is carried out to the characteristic point in the image to be matched and the original image using Euclidean distance, And determine whether the match point is wrong with the size of the ratio between time nearly Euclidean distance by the nearest Euclidean distance of match point, if The match point is wrong, then rejects the match point.
Alternatively, it is assumed that the original image feature point set is combined into e, and the characteristics of image point set to be matched is f, then Euclidean Distance D calculation formula is as follows：
In formula：D represents the Euclidean distance between original image and image to be matched, e1Represent first characteristic point of original image, f1 Represent first characteristic point of image to be matched, e2Represent second characteristic point of original image, f2Represent the second of image to be matched Individual characteristic point, eiRepresent the ith feature point of original image, fiRepresent the ith feature point of image to be matched；
Assuming that the nearest Euclidean distance of match point is Dmin, secondary nearly Euclidean distance is DsecIf
Then whether the match point is wrong, wherein, K is default threshold value, 0 ＜ K ＜ 1.
Alternatively, it is described that the characteristic point of input picture is detected in metric space, including：
Step A1：Gray proces are carried out to the input picture, obtain gray level image；
Step A2：Convolution behaviour is carried out to the integral image of the gray level image by the different Gaussian filter of size Make；
The position of characteristic point is detected by the Hessian matrixes in rapid robust feature SURF algorithm；Specifically, assume Certain point on the gray level image is I=(a, b), then the Hessian matrixes of the point are designated as：H(I,σ)：
In formula：Laa(I, σ) represents Gauss second dervativeWith gray level image I convolution, Lab(I, σ) represents Gauss Second dervativeWith gray level image I convolution, Lbb(I, σ) represents Gauss second dervativeWith gray level image I's Convolution, I represent gray level image, and σ represents the scale factor of second order Gauss wave filter；G (σ) represents the Gaussian function of gray level image；
Or by the different box filter of size come the integration instead of Gaussian filter to the gray level image Image carries out convolution operation；Wherein, the integral image of the gray level image is entered by size different box filter The approximate Hessian determinants of a matrix that row convolution operation obtains are as follows：
Det (H)=MaaMbb-(0.9Mab)2
In formula：Det (H) represents approximate Hessian determinants of a matrix, MaaBox filter and image I convolution are represented, Approximate Laa(I,σ)；MbbRepresent box filter and image I convolution, approximate Lbb(I,σ)；MabRepresent box filter and image I convolution, approximate Lab(I,σ)；
Step A3：4 × 4 × 4 local neighborhood is found in metric space, by each 16 points of yardstick above and below comparing and together 8 point of layer totally 47 points, to determine extreme point, and using three-dimensional quadratic fit method come precise positioning feature point.
Alternatively, the principal direction for determining the characteristic point, including：
Centered on characteristic point, the border circular areas that a radius is 6 σ is established, is with central angleThe anglec of rotation is's Sector region rotates a circle around the characteristic point, obtains the Haar small echos of the characteristic point in the border circular areas in all sectors Sum is responded, takes most long small echo to respond and is used as characteristic point principal direction；Wherein σ represents feature point scale.
Alternatively, the block for determining zoning centered on the characteristic point, the zoning being divided into n × n Region, including：
Centered on the characteristic point, the square area that length is 20 σ draws the zoning as zoning It is divided into 4 × 4 block region, calculates the both horizontally and vertically small echo response sum of 5 × 5 sample points in each block region, and The absolute value of small echo response is summed, obtains 64 dimensional feature description vectors of the input picture.
Alternatively, it is described that the feature description vectors of the input picture are carried out by dimensionality reduction by principal component analysis PCA algorithms Processing, obtains dimensionality reduction feature description vectors；Including：
Assuming that the eigenvectors matrix extracted by rapid robust feature SURF algorithm is X, wherein：
X={ X1,X2,…Xi…Xn}
In formula：X1Represent the 1st characteristic point matrix, X2Represent the 2nd characteristic point matrix, XiRepresent ith feature point square Battle array, XnN-th of characteristic point matrix is represented, n is characterized sum a little；
Order
In formula：C represents covariance matrix,Represent characteristic point matrix average；
Obtain eigenmatrix ν corresponding to covariance matrix CiAnd μi；
Eigenmatrix is subjected to descending arrangement according to the size of characteristic value, 8 dimensional vectors are as PCA before taking each eigenmatrix Dimensionality reduction matrix.
Alternatively, it is described to generate Hash sequence according to the dimensionality reduction feature description vectors, including：
Each row of n characteristic point matrix are overlapped, form feature Hash sequence, and by the feature Hash sequence Assemble initial Hash sequence；
Dimensionality reduction is carried out to the initial Hash sequence, obtains final Hash sequence.
Compared with prior art, the present invention has following beneficial effect：
The image Hash sequence generating method provided by the invention combined based on SURF and PCA, by extending three dimensions The method of middle comparison point number, counted out to reduce low quality feature, improve image recognition rate；Using PCA algorithms to characteristic point Dimensionality reduction is carried out, the data after dimensionality reduction are capable of the principal character of representative image, so as to effectively lift the discrimination of image, in protrusion There is good effect in terms of otherness, the difference for giving expression to two images can be understood.
Brief description of the drawings
The detailed description made by reading with reference to the following drawings to non-limiting example, further feature of the invention, Objects and advantages will become more apparent upon：
Fig. 1 is that Fig. 1 is the image Hash sequence generating method combined based on SURF and PCA that the embodiment of the present invention one provides Flow chart；
Fig. 2 is SURF extreme point comparison schematic diagrams；
Fig. 3 is characterized a principal direction schematic diagram；
Fig. 4 is SURF feature point description vector schematic diagrames；
Fig. 5 is that application provided by the invention carries out image based on the image Hash sequence generating method that SURF and PCA is combined The method flow diagram of identification；
Fig. 6 is the discrimination contrast signal of the method in the first width artwork application present invention and the hash method based on NMF Figure；
Fig. 7 is the discrimination contrast signal of the method in the second width artwork application present invention and the hash method based on NMF Figure.
Embodiment
With reference to specific embodiment, the present invention is described in detail.Following examples will be helpful to the technology of this area Personnel further understand the present invention, but the invention is not limited in any way.It should be pointed out that the ordinary skill to this area For personnel, without departing from the inventive concept of the premise, some changes and improvements can also be made.These belong to the present invention Protection domain.
Fig. 1 is the stream for the image Hash sequence generating method combined based on SURF and PCA that the embodiment of the present invention one provides Cheng Tu, as shown in figure 1, the method for the present embodiment includes：
S101, the characteristic point of input picture is detected in metric space；
Alternatively, step S101, can include：
Step A1：Gray proces are carried out to the input picture, obtain gray level image；
Step A2：Convolution behaviour is carried out to the integral image of the gray level image by the different Gaussian filter of size Make；
The position of characteristic point is detected by the Hessian matrixes in rapid robust feature SURF algorithm；Specifically, assume Certain point on the gray level image is I=(a, b), then the Hessian matrixes of the point are designated as：H(I,σ)：
In formula：Laa(I, σ) represents Gauss second dervativeWith gray level image I convolution, Lab(I, σ) represents Gauss Second dervativeWith gray level image I convolution, Lbb(I, σ) represents Gauss second dervativeWith gray level image I's Convolution, I represent gray level image, and σ represents the scale factor of second order Gauss wave filter；G (σ) represents the Gaussian function of gray level image；
Or by the different box filter of size come the integration instead of Gaussian filter to the gray level image Image carries out convolution operation；Wherein, the integral image of the gray level image is entered by size different box filter The approximate Hessian determinants of a matrix that row convolution operation obtains are as follows：
Det (H)=MaaMbb-(0.9Mab)2
In formula：Det (H) represents approximate Hessian determinants of a matrix, MaaBox filter and image I convolution are represented, Approximate Laa(I,σ)；MbbRepresent box filter and image I convolution, approximate Lbb(I,σ)；MabRepresent box filter and image I convolution, approximate Lab(I,σ)；
Step A3：4 × 4 × 4 local neighborhood is found in metric space, by each 16 points of yardstick above and below comparing and together 8 point of layer totally 47 points, to determine extreme point, and using three-dimensional quadratic fit method come precise positioning feature point.
In the present embodiment, the input picture includes：Image and original image to be matched；As shown in Fig. 2 Fig. 2 is SURF Extreme point comparison schematic diagram.In view of the situation that the characteristic point quantity extracted is excessive, matching is inaccurate, the present embodiment passes through expansion The method for opening up comparison point number in three dimensions, to reduce low quality characteristic point.Specifically 3 × 3 × 3 local neighborhood is improved Into 4 × 4 × 4 neighborhoods, by relatively each 16 points of yardstick and with 8 point of layer totally 47 points up and down, to determine extreme point, Ke Yijin One step improves the quality of characteristic point.Finally with three-dimensional quadratic fit method come precise positioning feature point.
S102, the principal direction for determining the characteristic point.
In the present embodiment, as shown in figure 3, Fig. 3 is characterized a principal direction schematic diagram；Centered on characteristic point, one is established Radius is 6 σ border circular areas, is with central angleThe anglec of rotation isSector region around the characteristic point rotation one In week, the Haar small echos response sum of the characteristic point in the border circular areas in all sectors is obtained, takes most long small echo to respond and makees It is characterized a principal direction；Wherein σ represents feature point scale.
S103, centered on the characteristic point, determine zoning, the zoning be divided into n × n block region.
In the present embodiment, centered on the characteristic point, length be 20 σ square area as zoning, by institute The block region that zoning is divided into 4 × 4 is stated, calculates the both horizontally and vertically small echo of 5 × 5 sample points in each block region Sum is responded, and the absolute value of small echo response is summed, obtains 64 dimensional feature description vectors of the input picture.Specifically , it is assumed that dxResponded for the Haar small echos of horizontal direction, dyResponded for the Haar small echos of vertical direction, respectively to each piece of region In 25 sample points dx, dy, | dx|, | dy| summed, obtain the 4 dimension description vectors V in each piece of region, wherein：V=(∑s dx,∑dy,∑|dx|,∑|dy|)；So each characteristic point has 16 × 4=64 dimension description vectors.As shown in figure 4, Fig. 4 is SURF feature point description vector schematic diagrames.The characteristic vector size of SURF algorithm extraction is n × 64 (n is characterized a number).It is right Than the SIFT feature vector of 128 dimensions, SURF reduces half in terms of dimension, and complexity reduces.
S104, multiple sample points in each block region of extraction both horizontally and vertically on small echo response sum, and The absolute value of small echo response is summed, obtains the feature description vectors of the input picture.
S105, the feature description vectors of the input picture are carried out by dimension-reduction treatment by principal component analysis PCA algorithms, obtained To dimensionality reduction feature description vectors.
In the present embodiment, it is assumed that the eigenvectors matrix extracted by rapid robust feature SURF algorithm is X, wherein：
X={ X1,X2,…Xi…Xn}
In formula：X1Represent the 1st characteristic point matrix, X2Represent the 2nd characteristic point matrix, XiRepresent ith feature point square Battle array, XnN-th of characteristic point matrix is represented, n is characterized sum a little.
Order
In formula：C represents covariance matrix,Represent characteristic point matrix average.
Obtain eigenmatrix ν corresponding to covariance matrix CiAnd μi；Eigenmatrix is subjected to descending according to the size of characteristic value Arrangement, takes dimensionality reduction matrix of 8 dimensional vectors as PCA before each eigenmatrix.
S106, according to the dimensionality reduction feature description vectors generate Hash sequence.
In the present embodiment, each row of n characteristic point matrix are overlapped, form feature Hash sequence, and by described in Feature Hash arrangement set is into initial Hash sequence；Dimensionality reduction is carried out to the initial Hash sequence, obtains final Hash sequence.
Alternatively, characteristic point n takes 30, and each row (PCA base) of 30 eigenmatrixes are overlapped, and forms feature and breathes out Uncommon sequences hi, then the Hash arrangement set of this 30 characteristic points into Hash sequences h.Then dimensionality reduction is carried out to h with dimensionality reduction matrix, Hash sequence to the end is obtained, length is n × 8.
Alternatively, the feature description vectors of the input picture are being carried out at dimensionality reduction by principal component analysis PCA algorithms Before reason, in addition to：
Feature similarity measure is carried out to the characteristic point in the image to be matched and the original image using Euclidean distance, And determine whether the match point is wrong with the size of the ratio between time nearly Euclidean distance by the nearest Euclidean distance of match point, if The match point is wrong, then rejects the match point.The characteristic matching point quantity that SURF algorithm obtains is a lot, and SURF algorithm is in itself Just include and reject Mismatching point the step, for improving the precision of algorithm.Therefore dimensionality reduction being carried out in this step can obtain Preferably effect.
Specifically, assuming that the original image feature point set is combined into e, the characteristics of image point set to be matched is f, then Euclidean Distance D calculation formula is as follows：
In formula：D represents the Euclidean distance between original image and image to be matched, e1Represent first characteristic point of original image, f1 Represent first characteristic point of image to be matched, e2Represent second characteristic point of original image, f2Represent the second of image to be matched Individual characteristic point, eiRepresent the ith feature point of original image, fiRepresent the ith feature point of image to be matched；Specifically, characteristic point Quantity is a lot, and characteristic point quantity i does not have specific scope in early stage, when generating hash sequences, can lock some scope, example Such as it is arranged between 30-60.
Assuming that the nearest Euclidean distance of match point is Dmin, secondary nearly Euclidean distance is DsecIf
Then whether the match point is wrong, wherein, K is default threshold value, 0 ＜ K ＜ 1.
Fig. 5 is that application provided by the invention carries out image based on the image Hash sequence generating method that SURF and PCA is combined The method flow diagram of identification, matching is generated using the above-mentioned image Hash sequence generating method combined based on SURF and PCA The Hash sequence of image and original image；The Hash sequence of the matching image and original image is compared, obtains Hash performance detection knot Fruit.
Specifically, general features point number n takes 30~60, n takes 30 in the present invention.Hash is calculated using L2 norms Sequence similarity
Wherein, h1And h (i)2(i) h is represented1And h2I-th of element, a threshold value can be set, if P is less than the threshold value, Represent that image vision content is similar, and P is smaller, and image is more similar.
In order to evaluate the hash method performance invented herein, following experiment is carried out, using University of Washington's view data Image in storehouse (Ground Truth Database) is as experimental subjects.Experiment includes the idiographic flow of this paper inventive methods, The hash method based on NMF and the hash method invented herein are also have chosen to contrast simultaneously.Present invention experiment is to use MATLAB written in code, the personal notebook electricity that experimental situation is Intel (R) Core (TM) i7 CPU@2.60GHz, 8GB RAM Brain.
45 width images are selected from image data base, and original image is all kinds of to original image progress according to the parameter in table 1 Geometric attack, the 45 width to be matched image similar with original image can be obtained.Again 45 images to be matched respectively with sending out herein Bright hash method obtains the discrimination contrast of table 2 compared with the hash method based on NMF.
Table 1
Table 2
First width artwork is produced into 45 to be matched images similar to artwork by the operation of table 1, uses the inventive method Image recognition rate, discrimination comparing result as shown in Figure 6 are obtained with the hash method based on NMF.Similarly to the second accompanying drawing The operation for carrying out table 1 produces 45 to be matched images similar to artwork, can obtain the discrimination comparing result shown in Fig. 7. It can be seen that this paper inventive methods are highly resistant to all kinds of geometric transformations (change of scale, shearing, rotation, compression etc.), and have good Recognition effect, performance is apparently higher than based on (Non-negative Matrix Factorization, NMF) Non-negative Matrix Factorization Hash method.
The specific embodiment of the present invention is described above.It is to be appreciated that the invention is not limited in above-mentioned Particular implementation, those skilled in the art can make a variety of changes or change within the scope of the claims, this not shadow Ring the substantive content of the present invention.In the case where not conflicting, the feature in embodiments herein and embodiment can any phase Mutually combination.

Claims (8)

  1. A kind of 1. image Hash sequence generating method combined based on SURF and PCA, it is characterised in that including：
    The characteristic point of input picture is detected in metric space；The input picture includes：Image to be matched and original Image；
    Determine the principal direction of the characteristic point；
    Centered on the characteristic point, zoning is determined, the zoning is divided into n × n block region；
    Extract multiple sample points in each block region both horizontally and vertically on small echo response sum, and to the small echo The absolute value summation of response, obtains the feature description vectors of the input picture；
    The feature description vectors of the input picture are carried out by dimension-reduction treatment by principal component analysis PCA algorithms, obtain dimensionality reduction spy Levy description vectors；
    Hash sequence is generated according to the dimensionality reduction feature description vectors.
  2. 2. the image Hash sequence generating method according to claim 1 combined based on SURF and PCA, it is characterised in that Before the feature description vectors of the input picture are carried out into dimension-reduction treatment by principal component analysis PCA algorithms, in addition to：
    Feature similarity measure is carried out to the characteristic point in the image to be matched and the original image using Euclidean distance, and led to The size of the nearest Euclidean distance of overmatching point and the ratio between time nearly Euclidean distance determines whether the match point is wrong, if described Match point is wrong, then rejects the match point.
  3. 3. the image Hash sequence generating method according to claim 2 combined based on SURF and PCA, it is characterised in that Assuming that the original image feature point set is combined into e, the characteristics of image point set to be matched is f, then Euclidean distance D calculation formula It is as follows：
    <mrow> <mi>D</mi> <mo>=</mo> <msqrt> <mrow> <msup> <mrow> <mo>(</mo> <msub> <mi>e</mi> <mn>1</mn> </msub> <mo>-</mo> <msub> <mi>f</mi> <mn>1</mn> </msub> <mo>)</mo> </mrow> <mn>2</mn> </msup> <mo>+</mo> <msup> <mrow> <mo>(</mo> <msub> <mi>e</mi> <mn>2</mn> </msub> <mo>-</mo> <msub> <mi>f</mi> <mn>2</mn> </msub> <mo>)</mo> </mrow> <mn>2</mn> </msup> <mo>+</mo> <mn>...</mn> <mo>+</mo> <msup> <mrow> <mo>(</mo> <msub> <mi>e</mi> <mi>i</mi> </msub> <mo>-</mo> <msub> <mi>f</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> <mn>2</mn> </msup> </mrow> </msqrt> </mrow>
    In formula：D represents the Euclidean distance between original image and image to be matched, e1Represent first characteristic point of original image, f1Represent First characteristic point of image to be matched, e2Represent second characteristic point of original image, f2Represent second spy of image to be matched Levy point, eiRepresent the ith feature point of original image, fiRepresent the ith feature point of image to be matched；
    Assuming that the nearest Euclidean distance of match point is Dmin, secondary nearly Euclidean distance is DsecIf
    Then whether the match point is wrong, wherein, K is default threshold value, 0 ＜ K ＜ 1.
  4. 4. the image Hash sequence generating method according to claim 1 combined based on SURF and PCA, it is characterised in that It is described that the characteristic point of input picture is detected in metric space, including：
    Step A1：Gray proces are carried out to the input picture, obtain gray level image；
    Step A2：Convolution operation is carried out to the integral image of the gray level image by the different Gaussian filter of size；
    The position of characteristic point is detected by the Hessian matrixes in rapid robust feature SURF algorithm；Specifically, described in assuming Certain point on gray level image is I=(a, b), then the Hessian matrixes of the point are designated as：H(I,σ)：
    <mrow> <mi>H</mi> <mrow> <mo>(</mo> <mi>I</mi> <mo>,</mo> <mi>&amp;sigma;</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfenced open = "[" close = "]"> <mtable> <mtr> <mtd> <mrow> <msub> <mi>L</mi> <mrow> <mi>a</mi> <mi>a</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>I</mi> <mo>,</mo> <mi>&amp;sigma;</mi> <mo>)</mo> </mrow> </mrow> </mtd> <mtd> <mrow> <msub> <mi>L</mi> <mrow> <mi>a</mi> <mi>b</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>I</mi> <mo>,</mo> <mi>&amp;sigma;</mi> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msub> <mi>L</mi> <mrow> <mi>a</mi> <mi>b</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>I</mi> <mo>,</mo> <mi>&amp;sigma;</mi> <mo>)</mo> </mrow> </mrow> </mtd> <mtd> <mrow> <msub> <mi>L</mi> <mrow> <mi>b</mi> <mi>b</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>I</mi> <mo>,</mo> <mi>&amp;sigma;</mi> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> </mtable> </mfenced> </mrow>
    <mrow> <mi>g</mi> <mrow> <mo>(</mo> <mi>&amp;sigma;</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <mn>2</mn> <msup> <mi>&amp;pi;&amp;sigma;</mi> <mn>2</mn> </msup> </mrow> </mfrac> <mi>exp</mi> <mrow> <mo>(</mo> <mo>-</mo> <mo>(</mo> <mrow> <msup> <mi>a</mi> <mn>2</mn> </msup> <mo>+</mo> <msup> <mi>b</mi> <mn>2</mn> </msup> </mrow> <mo>)</mo> <mo>/</mo> <mn>2</mn> <msup> <mi>&amp;sigma;</mi> <mn>2</mn> </msup> <mo>)</mo> </mrow> </mrow>
    In formula：Laa(I, σ) represents Gauss second dervativeWith gray level image I convolution, Lab(I, σ) represents Gauss second order DerivativeWith gray level image I convolution, Lbb(I, σ) represents Gauss second dervativeWith gray level image I convolution, I represents gray level image, and σ represents the scale factor of second order Gauss wave filter；G (σ) represents the Gaussian function of gray level image；
    Or by the different box filter of size come the integral image instead of Gaussian filter to the gray level image Carry out convolution operation；Wherein, the integral image of the gray level image is rolled up by size different box filter The approximate Hessian determinants of a matrix that product operation obtains are as follows：
    Det (H)=MaaMbb-(0.9Mab)2
    In formula：Det (H) represents approximate Hessian determinants of a matrix, MaaBox filter and image I convolution are represented, it is approximate Laa(I,σ)；MbbRepresent box filter and image I convolution, approximate Lbb(I,σ)；MabRepresent box filter and image I's Convolution, approximate Lab(I,σ)；
    Step A3：4 × 4 × 4 local neighborhood is found in metric space, by comparing each 16 points of yardstick up and down and with layer 8 Individual point totally 47 points, to determine extreme point, and using three-dimensional quadratic fit method come precise positioning feature point.
  5. 5. the image Hash sequence generating method according to claim 1 combined based on SURF and PCA, it is characterised in that The principal direction for determining the characteristic point, including：
    Centered on characteristic point, the border circular areas that a radius is 6 σ is established, is with central angleThe anglec of rotation isSector Region rotates a circle around the characteristic point, obtains the Haar small echos response of the characteristic point in the border circular areas in all sectors Sum, take most long small echo to respond and be used as characteristic point principal direction；Wherein σ represents feature point scale.
  6. 6. the image Hash sequence generating method according to claim 1 combined based on SURF and PCA, it is characterised in that It is described to determine zoning centered on the characteristic point, the zoning is divided into n × n block region, including：
    Centered on the characteristic point, length be 20 σ square area as zoning, the zoning is divided into 4 × 4 block region, the both horizontally and vertically small echo response sum of 5 × 5 sample points in each block region is calculated, and to institute The absolute value summation of small echo response is stated, obtains 64 dimensional feature description vectors of the input picture.
  7. 7. the image Hash sequence generating method combined based on SURF and PCA according to any one of claim 1-6, its It is characterised by, it is described that the feature description vectors of the input picture are carried out by dimension-reduction treatment by principal component analysis PCA algorithms, obtain To dimensionality reduction feature description vectors；Including：
    Assuming that the eigenvectors matrix extracted by rapid robust feature SURF algorithm is X, wherein：
    X={ X1,X2,…Xi…Xn}
    In formula：X1Represent the 1st characteristic point matrix, X2Represent the 2nd characteristic point matrix, XiRepresent ith feature dot matrix, XnTable Show n-th of characteristic point matrix, n is characterized sum a little；
    Order
    In formula：C represents covariance matrix,Represent characteristic point matrix average；
    Obtain eigenmatrix ν corresponding to covariance matrix CiAnd μi；
    Eigenmatrix is subjected to descending arrangement according to the size of characteristic value, takes drop of 8 dimensional vectors as PCA before each eigenmatrix Tie up matrix.
  8. 8. the image Hash sequence generating method according to claim 7 combined based on SURF and PCA, it is characterised in that It is described to generate Hash sequence according to the dimensionality reduction feature description vectors, including：
    Each row of n characteristic point matrix are overlapped, form feature Hash sequence, and by the feature Hash arrangement set Into initial Hash sequence；
    Dimensionality reduction is carried out to the initial Hash sequence, obtains final Hash sequence.
CN201710652338.9A 2017-08-02 2017-08-02 The image Hash sequence generating method combined based on SURF and PCA Pending CN107423768A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710652338.9A CN107423768A (en) 2017-08-02 2017-08-02 The image Hash sequence generating method combined based on SURF and PCA

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710652338.9A CN107423768A (en) 2017-08-02 2017-08-02 The image Hash sequence generating method combined based on SURF and PCA

Publications (1)

Publication Number Publication Date
CN107423768A true CN107423768A (en) 2017-12-01

Family

ID=60437195

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710652338.9A Pending CN107423768A (en) 2017-08-02 2017-08-02 The image Hash sequence generating method combined based on SURF and PCA

Country Status (1)

Country Link
CN (1) CN107423768A (en)

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108038488A (en) * 2017-12-06 2018-05-15 河海大学常州校区 The robustness image hash method mixed based on SIFT and LBP
CN108960280A (en) * 2018-05-21 2018-12-07 北京中科闻歌科技股份有限公司 A kind of picture similarity detection method and system
CN109559273A (en) * 2018-11-23 2019-04-02 河北工业大学 A kind of quick joining method towards vehicle base map picture
CN111104872A (en) * 2019-11-29 2020-05-05 兰州交通大学 GF-2 image integrity authentication method applying SIFT and SVD perceptual hashing
CN111340109A (en) * 2020-02-25 2020-06-26 深圳市景阳科技股份有限公司 Image matching method, device, equipment and storage medium
CN111695507A (en) * 2020-06-12 2020-09-22 桂林电子科技大学 Static gesture recognition method based on improved VGGNet network and PCA
CN111968027A (en) * 2020-08-20 2020-11-20 海南大学 Robust color image zero watermarking method based on SURF and DCT features
CN112734872A (en) * 2020-12-30 2021-04-30 天津工业大学 Fluorescence molecule tomography method and system based on multi-wavelength concurrent reconstruction
CN113408608A (en) * 2021-06-17 2021-09-17 四川大学 Similarity analysis method for similar copper coins based on computer vision
CN114265427A (en) * 2021-12-06 2022-04-01 江苏方天电力技术有限公司 Inspection unmanned aerial vehicle auxiliary navigation system and method based on infrared image matching
CN114758162A (en) * 2022-06-14 2022-07-15 北京市农林科学院信息技术研究中心 Commodity anti-counterfeiting identification method and device, electronic equipment and storage medium

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110264676A1 (en) * 2010-04-26 2011-10-27 Adi Belan Method and system for providing the download of transcoded files
CN104657383A (en) * 2013-11-22 2015-05-27 华中科技大学 Repeated video detection method and system based on correlation properties
CN105389593A (en) * 2015-11-16 2016-03-09 上海交通大学 Image object recognition method based on SURF

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110264676A1 (en) * 2010-04-26 2011-10-27 Adi Belan Method and system for providing the download of transcoded files
CN104657383A (en) * 2013-11-22 2015-05-27 华中科技大学 Repeated video detection method and system based on correlation properties
CN105389593A (en) * 2015-11-16 2016-03-09 上海交通大学 Image object recognition method based on SURF

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
孙锐等: "基于SIFT和PCA的图像感知哈希方法", 《电路与系统学报》 *
张珊: "基于PCA-SURF的单目视觉SLAM算法", 《万方数据知识服务平台》 *

Cited By (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108038488A (en) * 2017-12-06 2018-05-15 河海大学常州校区 The robustness image hash method mixed based on SIFT and LBP
CN108038488B (en) * 2017-12-06 2021-09-17 河海大学常州校区 Robustness image hashing method based on SIFT and LBP mixing
CN108960280A (en) * 2018-05-21 2018-12-07 北京中科闻歌科技股份有限公司 A kind of picture similarity detection method and system
CN108960280B (en) * 2018-05-21 2020-07-24 北京中科闻歌科技股份有限公司 Picture similarity detection method and system
CN109559273A (en) * 2018-11-23 2019-04-02 河北工业大学 A kind of quick joining method towards vehicle base map picture
CN111104872A (en) * 2019-11-29 2020-05-05 兰州交通大学 GF-2 image integrity authentication method applying SIFT and SVD perceptual hashing
CN111340109A (en) * 2020-02-25 2020-06-26 深圳市景阳科技股份有限公司 Image matching method, device, equipment and storage medium
CN111340109B (en) * 2020-02-25 2024-01-26 深圳市景阳科技股份有限公司 Image matching method, device, equipment and storage medium
CN111695507B (en) * 2020-06-12 2022-08-16 桂林电子科技大学 Static gesture recognition method based on improved VGGNet network and PCA
CN111695507A (en) * 2020-06-12 2020-09-22 桂林电子科技大学 Static gesture recognition method based on improved VGGNet network and PCA
CN111968027A (en) * 2020-08-20 2020-11-20 海南大学 Robust color image zero watermarking method based on SURF and DCT features
CN111968027B (en) * 2020-08-20 2022-11-01 海南大学 Robust color image zero watermarking method based on SURF and DCT features
CN112734872A (en) * 2020-12-30 2021-04-30 天津工业大学 Fluorescence molecule tomography method and system based on multi-wavelength concurrent reconstruction
CN112734872B (en) * 2020-12-30 2022-10-04 天津工业大学 Fluorescence molecule tomography method and system based on multi-wavelength concurrent reconstruction
CN113408608A (en) * 2021-06-17 2021-09-17 四川大学 Similarity analysis method for similar copper coins based on computer vision
CN114265427A (en) * 2021-12-06 2022-04-01 江苏方天电力技术有限公司 Inspection unmanned aerial vehicle auxiliary navigation system and method based on infrared image matching
CN114265427B (en) * 2021-12-06 2024-02-02 江苏方天电力技术有限公司 Inspection unmanned aerial vehicle auxiliary navigation system and method based on infrared image matching
CN114758162A (en) * 2022-06-14 2022-07-15 北京市农林科学院信息技术研究中心 Commodity anti-counterfeiting identification method and device, electronic equipment and storage medium

Similar Documents

Publication Publication Date Title
CN107423768A (en) The image Hash sequence generating method combined based on SURF and PCA
Wang et al. An image copy-move forgery detection method based on SURF and PCET
Jing et al. A face and palmprint recognition approach based on discriminant DCT feature extraction
Lee Copy-move image forgery detection based on Gabor magnitude
Liu et al. Content-based image retrieval using color difference histogram
Ross et al. A hybrid fingerprint matcher
US20180268195A1 (en) Gabor cube feature selection-based classification method and system for hyperspectral remote sensing images
CN103116763B (en) A kind of living body faces detection method based on hsv color Spatial Statistical Character
Woodard et al. Finger surface as a biometric identifier
CN103927511B (en) image identification method based on difference feature description
Herdiyeni et al. Combination of morphological, local binary pattern variance and color moments features for indonesian medicinal plants identification
CN111126240B (en) Three-channel feature fusion face recognition method
CN103839042B (en) Face identification method and face identification system
CN102332084B (en) Identity identification method based on palm print and human face feature extraction
CN102663371B (en) Low-resolution face recognition method coupling gait characteristics
CN101079104A (en) Human face identification method based on information
Rida et al. Forensic shoe-print identification: a brief survey
Liu et al. Improved SIFT-based copy-move detection using BFSN clustering and CFA features
CN110147800A (en) Image duplication based on SIFT, which is pasted, distorts blind detection method
Selvaraj et al. Enhanced copy–paste forgery detection in digital images using scale‐invariant feature transform
CN100416592C (en) Human face automatic identifying method based on data flow shape
CN107832793A (en) The sorting technique and system of a kind of high spectrum image
Deshpande et al. Latent fingerprint identification system based on a local combination of minutiae feature points
Kumar et al. Discrete cosine transform features matching-based forgery mask detection for copy-move forged images
Muzaffer et al. Recent keypoint based copy move forgery detection techniques

Legal Events

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

Application publication date: 20171201