CN102881008B - Based on the anti-rotation image Hash method of annulus statistical nature - Google Patents

Abstract

The present invention relates to a kind of anti-rotation image Hash method based on annulus statistical nature, first by bilinear interpolation normalized images size; Image is processed again with gauss low frequency filter; If coloured image, it is transformed into YCbCr color space and is got brightness representation in components; Then Iamge Segmentation is become some annulus, extracts the statistics such as the average of each annulus, variance, measure of skewness and kurtosis as feature; Then to statistical nature normalization method; Calculate again each statistic average and in this, as reference, finally calculate the Euclidean distance of annulus statistical nature and fixed reference feature, all distance values of contacting are image Hash; When judging similarity, calculate the L2 norm of two Hash, if being less than setting threshold value, it is believed that the image of its correspondence is identical, otherwise thinks different images. The common digital processings such as image rotation, JPEG compression, noise jamming, brightness adjustment, contrast strengthen steadily and surely and are had good uniqueness by the present invention.

Description

Based on the anti-rotation image Hash method of annulus statistical nature

Technical field

The present invention relates to signal processing field and field of computer technology, particularly a kind of anti-rotation image Hash method based on annulus statistical nature.

Background technology

The appearance of various image procossing software powerful, simple to operate; such as Photoshop and ACDSee; the editing operation of digital picture is made to become more and more easier; cause distorting, the image impinge issue such as forgery day by day serious, image content copyright protection and integrated authentication become thing urgently to be resolved hurrily. Meanwhile, digital picture increases on a large scale and is badly in need of more efficiently technology, to realize the efficient retrieval to mass image data and management. In recent years, has there is the new technology of a kind of image Hash (Ha Xi) by name in digital media content process field, can be widely used in the aspects such as image retrieval, copy detection, content authentication, tampering detection, numeral watermark, picture quality evaluation, image index.

Image Hash is a string short and small character for token image or numeral sequence. Due to normal digital processing, such as JPEG compression and image enhaucament, the concrete expression data of image can be changed, but still keep vision content constant, therefore, although conventional cipher Hash function, such as MD5 and SHA-1 to, the information that inputs arbitrarily can be converted the character string that length is fixing, but because they are very responsive to input change, any 1 bit difference all can change output completely, extracts so image Hash can not be directly used in. For this reason, image Hash should reflect vision content, represents that data are insensitive to concrete. Usually, image Hash should meet two conditions: (1) robustness, namely that vision is similar image, and no matter it specifically represents that whether data are identical, and Hash should be identical or very close with very big probability; (2) uniqueness, namely the image of different content has completely different Hash.

Image Hash research origin is in the middle and later periods the 90's of last century, but after entering for 21st century, this aspect research just really causes the extensive attention of researchist, and obtains very fast development in recent years. Inherent constructing technology according to image Hash method is different, existing method can be roughly divided into following five classes: the method for (1) Corpus--based Method amount; (2) based on the method for constant relation; (3) method described based on coarse features; (4) the semantic feature of the low layer of image is utilized; (5) matrix decomposition technology is used.

Some digital processing multipair greatly of above-mentioned prior art has robustness, as JPEG compression, digital filtering, image contracting put, brightness adjustment, but generally that rotation process is fragile, the deficiency that Existence and uniquenss is poor.

Summary of the invention

It is an object of the invention to provide a kind of anti-rotation image Hash method based on annulus statistical nature, the method can extract Shandong rod Hash from digital picture, and steadily and surely and the common digital processings such as image rotation, JPEG compression, brightness adjustment are had good uniqueness.

The technical scheme realizing the object of the invention is:

A kind of anti-rotation image Hash method based on annulus statistical nature, first convert input picture to specified dimension with bilinear interpolation, image is processed again with gauss low frequency filter, if input picture is coloured image, then it is transformed into YCbCr color space and is got brightness representation in components, then Iamge Segmentation is become some annulus, extract the average of each annulus pixel, variance, the statistic such as measure of skewness and kurtosis is as feature, then statistical nature is normalized, calculate again each statistic average and in this, as fixed reference feature, finally calculate the Euclidean distance of each annulus statistical nature and fixed reference feature, all distance values of contacting are image Hash, the flow process extracting image Hash is as shown in Figure 1, when judging Hash similarity, calculate their L2 norm, if being less than setting threshold value, it is believed that the image of their correspondences is identical, otherwise thinks different images.

Based on the anti-rotation image Hash method of annulus statistical nature, concrete steps are as follows:

(1) dimensions: input picture is normalized into M �� M size with bilinear interpolation;

(2) Gassian low-pass filter: normalized images is carried out filtering with 3 �� 3 gauss low frequency filters, after note process, image is I;

(3) color space conversion: if input picture is not coloured image, so makes J=I and turns to execution (4) step; Otherwise image I being transformed into YCbCr color space and gets brightness representation in components, note brightness component is J;

(4) image annular segmentation: the N number of annulus being divided into area roughly equal J with N number of circle, by order from small to large to the radius numbering of circle, if r_i(1��i��N) is the radius of i-th circle, and so the radius of N number of circle isWhereinRepresent downward rounding operation, then calculate annulus average areaSo the radius obtaining the 1st circle isThe radius of jth circle is(j=2,3,4 ..., N-1); If the pixel value of the capable xth row of the y that p (x, y) (1��x��M, 1��y��M) is J, with (x_c,y_c) represent image centre coordinate, if M is even number, get x_c=M/2+0.5 and y_c=M/2+0.5, otherwise get x_c=(M+1)/2 and y_c(M+1)/2, calculate the Euclidean distance of p (x, y) to figure inconocenterThe pixel value of each annulus is obtained, i.e. R according to the relation of pixel distance value and circle radius₁={p(x,y)|d_x,y��r₁And R_n={p(x,y)|r_n-1<d_x,y��r_n(n=2,3 ..., N), wherein R₁And R_nIt is the 1st and n-th set of pixel value of annulus respectively;

(5) annulus statistical nature is extracted: calculate R_iThe average of (1��i��N), variance, measure of skewness and kurtosis also use m_i��v_i��s_iAnd k_iRepresent, so obtaining 4 vectors, i.e. m=[m₁,m₂,...,m_N]��v=[v₁,v₂,...,v_N]��s=[s₁,s₂,...,s_N] and k=[k₁,k₂,...,k_N];

(6) feature normalization: respectively to the element normalization method of m, v, s and k and result is designated as m⁽¹⁾��v⁽¹⁾��s⁽¹⁾And k⁽¹⁾, i.e. m_i ⁽¹⁾=(m_i-��_m)/��_m��v_i ⁽¹⁾=(v_i-��_v)/��_v��s_i ⁽¹⁾=(s_i-��_s)/��_sAnd k_i ⁽¹⁾=(k_i-��_k)/��_k(1��i��N), wherein ��_m����_v����_sAnd ��_kIt is the average of m, v, s and k respectively, ��_m����_v����_sAnd ��_kThe standard deviation being respectively them, m_i ⁽¹⁾��v_i ⁽¹⁾��s_i ⁽¹⁾And k_i ⁽¹⁾It is m respectively⁽¹⁾��v⁽¹⁾��s⁽¹⁾And k⁽¹⁾I-th element;

(7) characteristic distance is calculated: calculate m respectively⁽¹⁾��v⁽¹⁾��s⁽¹⁾And k⁽¹⁾Average and using them as with reference to feature, namely $a_m=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{m}_i^{\left(1\right)},$ $a_v=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{v}_i^{\left(1\right)},$ $a_s=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{s}_i^{\left(1\right)}$ With $a_k=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{k}_i^{\left(1\right)},$ Then the Euclidean distance of each annulus feature and fixed reference feature is calculated $h_i=\sqrt{{(m_i^{\left(1\right)}-a_m)}^2+{(v_i^{\left(1\right)}-a_v)}^2+{(s_i^{\left(1\right)}-a_s)}^2+{(k_i^{\left(1\right)}-a_k)}^2}(1\&le;i\&le;N),$ So obtaining image Hashh=[h₁,h₂,...,h_N];

(8) similarity judges: establish h⁽¹⁾And h⁽²⁾It is two image Hash, calculates their L2 normWherein h_i ⁽¹⁾And h_i ⁽²⁾It is h respectively⁽¹⁾And h⁽²⁾I-th element, if d be less than setting threshold value T, it is believed that h⁽¹⁾And h⁽²⁾Corresponding image is identical, otherwise thinks different images.

It is an advantage of the invention that: the present invention is compared with prior art, there is following apparent outstanding substantive distinguishing features and marked improvement: Iamge Segmentation is become a series of annulus by the present invention, extract the statistics such as the average of annulus pixel, variance, measure of skewness and kurtosis as feature, rotation process is had good robustness; Utilize the average of each statistic of annulus as with reference to feature, obtaining Hash value by calculating the Euclidean distance of annulus feature and fixed reference feature; The calculated amount of characteristic extraction procedure is little, and travelling speed is fast; The common digital processings such as image rotation, JPEG compression, brightness adjustment, contrast strengthen, noise jamming steadily and surely and are had good uniqueness by the image Hash generated.

Accompanying drawing explanation

Fig. 1 is the schema that the present invention extracts image Hash;

Fig. 2 is the piece image used in embodiment, and size is 760 �� 560;

Fig. 3 is the another piece image similar to Fig. 2 vision, and size is 760 �� 560;

Fig. 4 is the result schematic diagram that Fig. 2 carries out dimensions;

Fig. 5 is the result schematic diagram that Fig. 4 carries out Gassian low-pass filter;

Fig. 6 is the result schematic diagram of the brightness component getting Fig. 5;

Fig. 7 is the annular segmentation schematic diagram of Fig. 6;

The graphic representation of L2 norm in Fig. 8 embodiment.

Embodiment

Do to illustrate in detail to a preferred embodiment of the present invention below in conjunction with accompanying drawing, but protection scope of the present invention is not limited to following embodiment.

Embodiment:

Fig. 1 is the schema that the present invention extracts image Hash.

The present embodiment comprises robustness checking and uniqueness verifies two portions. Robustness checking is that the distance of the image Hash corresponding to scaling system 2 and Fig. 3 is to judge that whether they are for similar image, compared with Fig. 2, Fig. 3 experienced a series of digital processing, comprises image and rotates (turn counterclockwise 4 ��), JPEG compression (quality factor is 50), brightness adjustment (adjusting range is 50) and white Gaussian noise (average be 0, variance be 0.01). In following step, (1) ~ (7) are the steps of the Hash extracting Fig. 2, owing to extracting the step of Hash and the identical of Fig. 2 of Fig. 3, repeat no more here, and (8) are the distances of the Hash of scaling system 2 and Fig. 3, and concrete steps are as follows:

(1) dimensions: with bilinear interpolation, Fig. 2 being normalized into 512 �� 512 sizes, result is as shown in Figure 4;

(2) Gassian low-pass filter: Fig. 4 is carried out filtering by 3 �� 3 gauss low frequency filters being 1 by standard deviation, and result is as shown in Figure 5;

(3) color space conversion: Fig. 5 being transformed into YCbCr color space and gets brightness representation in components, result is as shown in Figure 6;

(4) image annular segmentation: 60 annulus that Fig. 6 is divided into area roughly equal, Fig. 7 is image annular segmentation schematic diagram, calculates the set of the pixel value of each annulus, obtains following result:

R₁={175,170,166,...,64,64},R₂={177,177,176,...,59,58},R₃={221,221,220,...,56,56},

R₄={215,213,212,...,155,156},R₅={216,212,207,...,176,176},R₆={200,199,198,..,177,177},

R₇={212,207,202,...,178,178},R₈={199,197,194,...,179,180},R₉={209,206,206,...,180,180},

R₁₀={218,213,210,...,180,181},R₁₁={171,171,172,...,181,181},R₁₂={162,162,163,...,181,181},

R₁₃={176,180,183,...,179,180},R₁₄={161,161,161,...,182,182},R₁₅={152,155,155,...,181,18},

R₁₆={81,82,84,..,182,181},R₁₇={73,72,70,...,181,182},R₁₈={138,136,134,...,180,180},

R₁₉={159,159,159,...,180,179},R₂₀={142,130,114,...,180,179},R₂₁={157,158,158,...,181,181},

R₂₂={104,101,102,...,180,180},R₂₃={109,84,80,...,182,182},R₂₄={113,128,145,...,180,182},

R₂₅={111,113,114,...,183,183},R₂₆={111,110,110,...,183,183},R₂₇={140,141,135,...,183,183},

R₂₈={112,110,112,..,185,184},R₂₉={152,141,131,...,188,187},R₃₀={133,144,160,..,188,187},

R₃₁={186,185,184,...,190,189},R₃₂={181,181,183,...,195,191},R₃₃={130,143,164,..,185,186},

R₃₄={82,80,78,...,185,186},R₃₅={55,61,65,...,186,186},R₃₆={67,67,66,...,186,187},

R₃₇={63,63,63,...,188,188},R₃₈={50,52,56,..,190,191},R₃₉={61,61,61,...,191,191},

R₄₀={42,43,43,...,194,193},R₄₁={48,47,47,...,191,191},R₄₂={47,42,37,...,193,193},

R₄₃={50,50,47,...,192,192},R₄₄={58,55,55,...,193,193},R₄₅={63,62,62,...,192,193},

R₄₆={61,60,60,...,192,192},R₄₇={55,59,62,...,193,193},R₄₈={58,57,57,...,194,194},

R₄₉={64,65,65,...,193,193},R₅₀={47,53,58,...,193,193},R₅₁={56,56,56,...,192,192},

R₅₂={57,57,57,...,194,193},R₅₃={54,54,54,..,194,194},R₅₄={47,48,47,...,194,194},

R₅₅={43,45,47,...,193,194},R₅₆={43,38,33,...,195,194},R₅₇={50,51,51,..,194,194},

R₅₈={53,56,58,...,196,196},R₅₉={59,58,56,...,195,195},R₆₀={57,58,58,...,197,196,195};

(5) annulus statistical nature is extracted: calculate the average of all annulus, variance, measure of skewness and kurtosis, obtain following vector:

M=[78.06, 107.74, 115.95, 118.68, 129.89, 137.64, 141.22, 143.89, 146.51, 148.75, 148.13, 149.18, 146.3, 146.03, 147.46, 148.42, 145.88, 143.10, 144.19, 141.32, 139.26, 136.72, 133.73, 132.51, 130.07, 128.39, 128.10, 128.82, 128.79, 127.73, 128.08, 129.61, 129.08, 128.38, 125.03, 123.74, 122.92, 121.25, 120.09, 119.66, 119.02, 118.70, 118.40, 118.04, 117.69, 117.22, 116.61, 115.78, 114.85, 114.54, 114.25, 113.15, 111.69, 110.51, 109.88, 110.17, 110.14, 109.77, 109.50, 109.34],

V=[1390.3, 3415.4, 3835, 4183.5, 3714.1, 3321, 3068.7, 2948.7, 2847.1, 2748.8, 2791.9, 2533, 2504.8, 2353, 2149.4, 1896.6, 1913.7, 1916.3, 1847.9, 1969, 1869.5, 1840.9, 1892.1, 1846.2, 1838.9, 1854.2, 1862.6, 1806.2, 1830.4, 1902.9, 2000.5, 2001.2, 2020.8, 2074.9, 2147.3, 2254.7, 2358, 2442.5, 2504.6, 2682.7, 2968.9, 3217.4, 3397, 3489.8, 3595.7, 3650.6, 3693.6, 3759.9, 3882.4, 3945.9, 3961.7, 4020, 4050.9, 4080.3, 4159.7, 4142.8, 4156.3, 4141.4, 4145.9, 4148.7],

S=[1.809, 0.2712, 0.033, 0.1106,-0.3761,-0.6863,-0.9192,-1.0882,-1.2114,-1.2658,-1.2043,-1.2177,-1.0457,-0.983,-0.8428,-0.6058,-0.4035,-0.2339,-0.3833,-0.3544,-0.2318,-0.2101,-0.1842,-0.1996,-0.1784,-0.1739,-0.139,-0.1612,-0.1929,-0.1421,-0.0864,-0.072,-0.0158, 0.0165, 0.0932, 0.1278, 0.1811, 0.2613, 0.3407, 0.3604, 0.3479, 0.3229, 0.2705, 0.2492, 0.2155, 0.1967, 0.2013, 0.2141, 0.2227, 0.2223, 0.2228, 0.2166, 0.2287, 0.2412, 0.2251, 0.2214, 0.2172, 0.2074, 0.2072, 0.2221],

K=[5.5342, 1.4404, 1.3019, 1.3397, 1.5195, 1.8737, 2.2301, 2.6136, 2.993, 3.2548, 3.1541, 3.2092, 2.8697, 2.8506, 2.8256, 2.5231, 2.201, 2.1202, 2.4412, 2.3255, 2.1163, 2.0097, 1.8738, 1.88, 1.792, 1.6918, 1.6377, 1.667, 1.6781, 1.6874, 1.7148, 1.7751, 1.7791, 1.8271, 1.8096, 1.7615, 1.7412, 1.7671, 1.8318, 1.8808, 1.8951, 1.8992, 1.9312, 1.9421, 1.9619, 1.9399, 1.9392, 1.9383, 1.9141, 1.9263, 1.951, 1.946, 1.9502, 1.9587, 1.9141, 1.8998, 1.8609, 1.8284, 1.8189, 1.8345],

(6) feature normalization: to the element normalization method of m, v, s and k, obtain following result:

m⁽¹⁾=[-3.344,-1.2806,-0.7098,-0.52, 0.2594, 0.7982, 1.0471, 1.2327, 1.4148, 1.5706, 1.5275, 1.6005, 1.4002, 1.3815, 1.4809, 1.5476, 1.371, 1.1778, 1.2536, 1.054, 0.9108, 0.7342, 0.5263, 0.4415, 0.2719, 0.1551, 0.1349, 0.185, 0.1829, 0.1092, 0.1335, 0.2399, 0.2031, 0.1544,-0.0785,-0.1682,-0.2252,-0.3413,-0.4219,-0.4518,-0.4963,-0.5186,-0.5394,-0.5645,-0.5888,-0.6215,-0.6639,-0.7216,-0.7862,-0.8078,-0.828,-0.9044,-1.0059,-1.088,-1.1318,-1.1116,-1.1137,-1.1394,-1.1582,-1.1693],

v⁽¹⁾=[-1.6311, 0.6321, 1.1011, 1.4905, 0.9659, 0.5266, 0.2446, 0.1105,-0.003,-0.1129,-0.0647,-0.3541,-0.3856,-0.5552,-0.7828,-1.0653,-1.0462,-1.0433,-1.1197,-0.9844,-1.0956,-1.1276,-1.0703,-1.1216,-1.1298,-1.1127,-1.1033,-1.1663,-1.1393,-1.0583,-0.9492,-0.9484,-0.9265,-0.866,-0.7851,-0.6651,-0.5496,-0.4552,-0.3858,-0.1868, 0.1331, 0.4108, 0.6115, 0.7153, 0.8336, 0.895, 0.943, 1.0171, 1.154, 1.225, 1.2427, 1.3078, 1.3424, 1.3752, 1.4639, 1.4451, 1.4602, 1.4435, 1.4485, 1.4517],

s⁽¹⁾=[3.5982, 0.713, 0.2661, 0.4117,-0.5014,-1.0834,-1.5203,-1.8374,-2.0685,-2.1706,-2.0552,-2.0803,-1.7577,-1.64,-1.377,-0.9323,-0.5528,-0.2346,-0.5149,-0.4607,-0.2307,-0.19,-0.1414,-0.1703,-0.1305,-0.122,-0.0566,-0.0982,-0.1577,-0.0624, 0.0421, 0.0691, 0.1746, 0.2352, 0.3791, 0.444, 0.544, 0.6945, 0.8434, 0.8804, 0.8569, 0.81, 0.7117, 0.6718, 0.6085, 0.5733, 0.5819, 0.6059, 0.622, 0.6213, 0.6222, 0.6106, 0.6333, 0.6567, 0.6265, 0.6196, 0.6117, 0.5933, 0.593, 0.6209],

k⁽¹⁾=[5.4793,-1.0143,-1.234,-1.1741,-0.8889,-0.327, 0.2383, 0.8466, 1.4484, 1.8637, 1.7039, 1.7913, 1.2528, 1.2225, 1.1829, 0.703, 0.1921, 0.064, 0.5731, 0.3896, 0.0578,-0.1113,-0.3269,-0.3171,-0.4566,-0.6156,-0.7014,-0.6549,-0.6373,-0.6226,-0.5791,-0.4834,-0.4771,-0.401,-0.4287,-0.505,-0.5372,-0.4961,-0.3935,-0.3158,-0.2931,-0.2866,-0.2358,-0.2185,-0.1871,-0.222,-0.2231,-0.2246,-0.263,-0.2436,-0.2044,-0.2124,-0.2057,-0.1922,-0.263,-0.2856,-0.3473,-0.3989,-0.414,-0.3892],

(7) characteristic distance is calculated: calculate m⁽¹⁾��v⁽¹⁾��s⁽¹⁾And k⁽¹⁾Average, obtain fixed reference feature: a_m=-0.0000017, a_v=-0.0000017��a_s=-0.0000067 and a_k=0.000005, calculate the Euclidean distance of each annulus feature and fixed reference feature $h_i=\sqrt{{(m_i^{\left(1\right)}-a_m)}^2+{(v_i^{\left(1\right)}-a_v)}^2+{(s_i^{\left(1\right)}-a_s)}^2+{(k_i^{\left(1\right)}-a_k)}^2}(1\&le;i\&le;60),$ The image Hash of Fig. 2 can be obtained:

h⁽¹⁾=[7.5374, 1.8912, 1.8193, 2.0100, 1.4289, 1.4816, 1.8773, 2.3716, 2.8945, 3.2656, 3.0764, 3.1974, 2.6016, 2.5300, 2.4701, 2.2121, 1.8212, 1.5921, 1.8490, 1.5633, 1.4445, 1.3635, 1.2447, 1.2580, 1.2553, 1.2869, 1.3155, 1.3539, 1.3276, 1.2343, 1.1207, 1.0934, 1.0760, 0.9949, 0.9747, 0.9606, 0.9681, 1.0257, 1.0923, 1.0554, 1.0413, 1.0844, 1.1077, 1.1530, 1.2029, 1.2511, 1.3109, 1.4046, 1.5511, 1.6120, 1.6306, 1.7165, 1.8048, 1.8823, 1.9712, 1.9467, 1.9665, 1.9731, 1.9906, 2.0029],

(8) similarity judges: the image Hash extracting Fig. 3 obtains:

h⁽²⁾=[6.7303, 2.1643, 1.9925, 2.0166, 1.4411, 1.3637, 1.7813, 2.3103, 2.9400, 3.3054, 3.2468, 3.3685, 3.1246, 2.7672, 2.9203, 2.5954, 1.8488, 1.7239, 2.0789, 1.7923, 1.4754, 1.4068, 1.2303, 1.1985, 1.1251, 1.2318, 1.2110, 1.1374, 1.1292, 1.1129, 1.1451, 1.1112, 1.1140, 1.0188, 1.0856, 1.1292, 1.1498, 1.1321, 1.1266, 1.1038, 0.9629, 0.9486, 0.9090, 1.0437, 1.1291, 1.2667, 1.3086, 1.3680, 1.4082, 1.4805, 1.5354, 1.6260, 1.7373, 1.8618, 1.9855, 1.9708, 2.0215, 2.0462, 2.0021, 1.9904], setting threshold value T=2.0, calculates h⁽¹⁾And h⁽²⁾L2 norm, obtain d=1.4249; Due to d < T, therefore think that Fig. 2 with Fig. 3 is the similar images of two width visions.

The test data of uniqueness checking is the coloured image that 200 width contents are different, size is 256 �� 256��2048 �� 1536, (M=512 is got with the present invention, N=60, the standard deviation of gauss low frequency filter is 1) extract the Hash of 200 width images and calculate L2 norm between Hash between two, obtain altogether 19900 distance values, wherein minor increment is 2.608, ultimate range is 15.3646, and distance average is 8.4142, and standard deviation is 1.9297. Fig. 8 is the graphic representation of L2 norm, and wherein X-coordinate is the index of often couple of image Hash, and ordinate zou is L2 norm. Finding from embodiment, the Hash of any two width different images, its distance is all greater than setting threshold value T=2.0, illustrates that the present invention has good uniqueness.

Claims (1)

1. based on an anti-rotation image Hash method for annulus statistical nature, comprising: dimensions: with bilinear interpolation, input picture is normalized into M �� M size; Filtering: after note process, image is I; Color space conversion: if I is not coloured image, so makes J=I and turns to and perform image annular segmentation, otherwise image I is transformed into YCbCr color space and extract light intensity level, and note brightness component is J; Image annular segmentation; Extract annulus statistical nature; Feature normalization: respectively to the element normalization method of m, v, s and k and result is designated as m⁽¹⁾��v⁽¹⁾��s⁽¹⁾And k⁽¹⁾, i.e. m_i ⁽¹⁾=(m_i-��_m)/d_m��v_i ⁽¹⁾=(v_i-��_v)/d_v��s_i ⁽¹⁾=(s_i-��_s)/d_sAnd k_i ⁽¹⁾=(k_i-��_k)/d_k, wherein 1��i��N, ��_m����_v����_sAnd ��_kIt is the average of m, v, s and k respectively, d_m��d_v��d_sAnd d_kThe standard deviation being respectively them, m_i ⁽¹⁾��v_i ⁽¹⁾��s_i ⁽¹⁾And k_i ⁽¹⁾It is m respectively⁽¹⁾��v⁽¹⁾��s⁽¹⁾And k⁽¹⁾I-th element; Calculate characteristic distance: calculate m respectively⁽¹⁾��v⁽¹⁾��s⁽¹⁾And k⁽¹⁾Average and using them as with reference to feature, namely $a_m=\frac{1}{N}\underset{i=1}{\overset{N}{\&Sigma;}}{m}_i^{\left(1\right)},a_v=\frac{1}{N}\underset{i=1}{\overset{N}{\&Sigma;}}{v}_i^{\left(1\right)},a_s=\frac{1}{N}\underset{i=1}{\overset{N}{\&Sigma;}}{s}_i^{\left(1\right)}$ WithThen the Euclidean distance of each annulus feature and fixed reference feature is calculated $h_i=\sqrt{{(m_i^{\left(1\right)}-a_m)}^2+{(v_i^{\left(1\right)}-a_v)}^2+{(s_i^{\left(1\right)}-a_s)}^2+{(k_i^{\left(1\right)}-a_k)}^2},$ Wherein 1��i��N, so obtaining image Hashh=[h₁,h₂,...,h_N]; Similarity judges; It is characterized in that:

(1) Gassian low-pass filter it is filtered into described in: with 3 �� 3 gauss low frequency filters, normalized images is carried out filtering;

(2) described image annular segmentation: the N number of annulus being divided into area roughly equal J with N number of circle, by order from small to large to the radius numbering of circle, if r_iBeing the radius of i-th circle, wherein, 1��i��N, so the radius of N number of circle isWhereinRepresent downward rounding operation, then calculate annulus average areaSo the radius obtaining the 1st circle isThe radius of jth circle isWherein, j=2,3,4 ..., N-1; If the pixel value of the capable xth row of the y that p (x, y) is J, wherein, 1��x��M and 1��y��M, with (x_c,y_c) represent image centre coordinate, if M is even number, get x_c=M/2+0.5 and y_c=M/2+0.5, otherwise get x_c=(M+1)/2 and y_c=(M+1)/2, calculate the Euclidean distance of p (x, y) to figure inconocenterThe pixel value of each annulus is obtained, i.e. R according to the relation of pixel distance value and circle radius₁=p (x, y) | d_x,y��r₁And R_n=p (x, y) | r_n-1<d_x,y��r_n, wherein, R₁And R_nIt is the 1st and n-th set of pixel value of annulus respectively, n=2,3 ..., N;

(3) described extraction annulus statistical nature: calculate R_iAverage, variance, measure of skewness and kurtosis and use m_i��v_i��s_iAnd k_iRepresenting, wherein, 1��i��N, so obtaining 4 vectors, i.e. m=[m₁,m₂,...,m_N], v=[v₁,v₂,...,v_N], s=[s₁,s₂,...,s_N] and k=[k₁,k₂,...,k_N];

(4) described similarity judges: establish h⁽¹⁾And h⁽²⁾It is two image Hash, calculates their L2 normWherein h_i ⁽¹⁾And h_i ⁽²⁾It is h respectively⁽¹⁾And h⁽²⁾I-th element, if d be less than setting threshold value T, it is believed that h⁽¹⁾And h⁽²⁾Corresponding image is identical, otherwise thinks different images.