CN101271521A - Human face recognition method based on anisotropic double-tree complex wavelet package transforms - Google Patents
Human face recognition method based on anisotropic double-tree complex wavelet package transforms Download PDFInfo
- Publication number
- CN101271521A CN101271521A CNA2008101063962A CN200810106396A CN101271521A CN 101271521 A CN101271521 A CN 101271521A CN A2008101063962 A CNA2008101063962 A CN A2008101063962A CN 200810106396 A CN200810106396 A CN 200810106396A CN 101271521 A CN101271521 A CN 101271521A
- Authority
- CN
- China
- Prior art keywords
- wavelet
- face
- band
- anisotropy
- range coefficient
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 29
- 230000001815 facial effect Effects 0.000 claims description 87
- 239000004744 fabric Substances 0.000 claims description 27
- 238000006243 chemical reaction Methods 0.000 claims description 24
- 239000000284 extract Substances 0.000 claims description 15
- 238000012545 processing Methods 0.000 claims description 12
- 238000010606 normalization Methods 0.000 claims description 8
- 230000008676 import Effects 0.000 claims description 4
- 238000013507 mapping Methods 0.000 claims description 4
- 230000009466 transformation Effects 0.000 abstract description 14
- 238000000605 extraction Methods 0.000 abstract description 7
- 230000008901 benefit Effects 0.000 abstract description 4
- 238000005516 engineering process Methods 0.000 abstract description 4
- 230000014509 gene expression Effects 0.000 description 6
- 238000005286 illumination Methods 0.000 description 5
- 230000003542 behavioural effect Effects 0.000 description 4
- 230000008859 change Effects 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 210000000887 face Anatomy 0.000 description 3
- 238000012935 Averaging Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000000844 transformation Methods 0.000 description 2
- 239000002537 cosmetic Substances 0.000 description 1
- 230000005021 gait Effects 0.000 description 1
- 230000013011 mating Effects 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 210000001938 protoplast Anatomy 0.000 description 1
- 210000001525 retina Anatomy 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
- 230000001755 vocal effect Effects 0.000 description 1
Images
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention relates to a human face recognition method based on the anisotropy dual-tree complex wavelet packet transformation and belongs to a mode recognition technology field. The method includes that: firstly, an average face is processed; the characteristics of the face represented by a wavelet amplitude coefficient is obtained through a characteristic extraction of an input face image; a weight coefficient is used for the weighting of every wavelet sub-band amplitude coefficient, the same treatment is implemented on every regular front gray face image in a regular face database, and a standard face characteristic database is obtained; the face characteristics with wavelet amplitude coefficient corresponding to the input face image to be recognized matches one-to-one with the face characteristics with wavelet amplitude coefficient corresponding to every face image in the regular face characteristic database, and the face with the maximum similarity in the regular face characteristic database is used as the result of the face recognition. The face recognition method based on the anisotropy dual-tree complex wavelet packet transformation has the advantages of high face recognition accuracy as well as low computational complexity.
Description
Technical field
The invention belongs to mode identification technology, be specifically related to a kind of face identification method based on the anisotropy double-tree complex wavelet package transforms.
Background technology
Identification is very important in the daily life of mankind nowadays society with checking.Utilize human that had, the physiological characteristic that can unique its identity of sign itself or the identification and the checking of the identity that behavioural characteristic is carried out the people, be called as living things feature recognition.These biological characteristics comprise physiological characteristic and behavioural characteristic.Physiological characteristic comprises people's face, fingerprint, iris, retina, palmmprint, hand shape, DNA, auricle shape etc.; Behavioural characteristic comprises person's handwriting, vocal print, gait of people etc.These physiological characteristics and behavioural characteristic satisfy people's different uniqueness and the stability that remains unchanged for a long period of time at least to a certain extent, can reflect someone's individual characteristic, and corresponding one by one with the identity of individuality, thereby can be used for verifying the true and false of individual identity.
Numerous living things feature recognition methods is each has something to recommend him on performances such as the security aspect identification and the checking, reliability.Wherein, in year, recognition of face has obtained the attention of a lot of researchists and application at nearest ten or twenty.Recognition of face has following characteristics:
1) recognition of face meets mankind itself's identification custom, is that the mankind are used for the most important means of identification mutually;
2) recognition of face can hidden operation, has original advantage aspect security monitoring;
3) recognition of face is contactless, and the collection of facial image or model is the property invaded not, is accepted by people easily;
When 4) carrying out recognition of face, the collection of image is convenient, and collecting device can use low and middle-grade CCD/CMOS cameras, and is cheap, has very strong practicality;
5) certain, recognition of face also has shortcomings such as poor stability, reliability view data mass discrepancy lower, that collect are bigger, and this mainly is because people's face can be with the very big change of change generation of conditions such as age, cosmetic, illumination, visual angle, distance.
When carrying out recognition of face, the first step will be carried out pre-service to the gray scale facial image that obtains, and obtains the gray scale facial image of rule.Fundamental purpose is the facial image that obtains better quality, thereby improves the accuracy of recognition of face.The pretreated flow process of facial image generally is divided into following steps as shown in Figure 1:
1) input facial image;
2) to the facial image of input, the position according to the eyes of people in the facial image with different facial image registrations, makes that the different position of people's face in image is identical; Facial image behind the registration cut out be same size;
3) facial image that carried out registration and cut out is carried out histogram equalization.
The gray scale facial image of 4) output rule.
In various face identification methods, it is a crucial step that face characteristic extracts.At present, a kind of face feature extraction method of main flow is to adopt the Gabor wavelet transformation.Use the Gabor wavelet transformation to carry out the face characteristic extraction and just be to use Gabor small echo kernel function that primitive man's face image pixel is carried out convolution algorithm one by one, obtain Gabor small echo range coefficient feature with different frequency yardstick and direction.Gabor small echo kernel function as shown in the formula:
In the formula, u, locations of pixels in the v presentation video,
The dimensions in frequency of representing different Gabor small echo kernel functions, θ represents the direction of different Gabor small echo kernel functions.The Gabor wavelet transformation has different dimensions in frequency and direction, has good performance on face characteristic extracts.Although the Gabor conversion has shown superior performance on face characteristic extracts, its maximum shortcoming is that computation complexity is too high, implements too inconvenient.For example, use 4 dimensions in frequency (promptly
Value is 1,2,3,4) 8 directions (be that the θ value is 0,
) the Gabor wavelet transformation carry out face characteristic when extracting, need carry out convolution algorithm one by one to all pixels in a 4 * 8=32 Gabor small echo kernel function and the width of cloth facial image, operand is very huge.And the number of the Gabor small echo range coefficient of the expression face characteristic that obtains is 32 times of original facial image size, also is very big next carrying out the operand of image when mating between any two.Simultaneously, the Gabor wavelet transformation is not can both obtain good effect under all conditions.
Dual-tree complex wavelet is transformed to face characteristic a kind of new method is provided.The dual-tree complex wavelet conversion adopts two real number wavelet transformations to realize, first real number wavelet transformation is corresponding to the real part of dual-tree complex wavelet conversion, and second real number wavelet transformation is corresponding to the imaginary part of dual-tree complex wavelet conversion.These two real number wavelet transformations use different bank of filters, and each bank of filters all satisfies complete reconstruction condition, and these two real number wavelet transformation Hilbert transforms each other make the dual-tree complex wavelet conversion constitute and resolve small echo.The excellent characteristic that the dual-tree complex wavelet conversion has comprises: approximate TIME SHIFT INVARIANCE, and reconstruct fully, the directional selectivity in the two-dimensional space (has-75 ° in the two-dimensional space,-45 ° ,-15 °, 15 °, 45 °, 6 directions such as 75 ° of grades) and low computational complexity (be O (n under the two-dimensional case
2)).
The dual-tree complex wavelet conversion can adopt the small echo lifting scheme to realize.The small echo lifting scheme comprises three steps: division, prediction and renewal.Step toward division is meant original signal is divided into two parts: soon signal is divided into two subclass by strange, the idol of its sequence number: even number sequence subset and odd number preface subclass.Prediction steps is the correlativity of utilizing between two subclass, predicts another with a subclass, for example uses the odd number sequence subset to predict the even number sequence subset, thereby obtains the detail signal part, and deposit in the even number sequence subset.Step of updating is meant that using detail signal is that the even number sequence subset upgrades the odd number sequence subset, obtains the general picture signal, and deposits in the odd number sequence subset.Like this, the lifting scheme of wavelet transformation has been broken down into several very simple basic steps, and computational complexity is lower.
Though dual-tree complex wavelet conversion operand is less, performance is not good on the extraction face characteristic.Extract iff simply the dual-tree complex wavelet conversion being applied to face characteristic, the performance of recognition of face is not high.At present, still nobody is used for recognition of face to the dual-tree complex wavelet conversion.
Summary of the invention
The objective of the invention is to overcome the weak point of prior art, a kind of face identification method of efficient, robust is provided, when reducing computational complexity greatly, guarantee that recognition of face still has very high accuracy.
Face identification method based on the anisotropy double-tree complex wavelet package transforms proposed by the invention comprises that processing, the face characteristic to average face extracts and classification judgement three parts, specifically comprises:
1) described processing to average face specifically may further comprise the steps:
10) the regular positive gray scale facial image of several (being generally more than tens width of cloth) different people in the regular face database is averaged obtain average face;
11) adopt the small echo lifting scheme to carry out J level (J is a positive integer) dual-tree complex wavelet conversion to the average face image and obtain wavelet sub-bands at different levels, described wavelet sub-band has 2 low frequency wavelet subbands and the 1st, 2 of J+1 level ..., each 6 high frequency wavelet subband of J level;
12) to wherein the 1st, 2, ..., J-1 level high frequency wavelet subband adopts the small echo lifting scheme to carry out the anisotropy wavelet package transforms (promptly to j level (j=1 according to variance criterion, 2, ..., J-1) the high frequency wavelet subband carries out J-j level anisotropy wavelet package transforms), obtain anisotropy wavelet package transforms structure and corresponding small echo range coefficient;
Above-mentioned steps 11), 12) be collectively referred to as the anisotropy double-tree complex wavelet package transforms;
13) the small echo range coefficient that obtains through the anisotropy double-tree complex wavelet package transforms is carried out normalization, the range coefficient in each wavelet sub-band evenly is normalized to 0 to 255 from small to large;
14) calculate the standard deviation of each wavelet sub-band range coefficient;
2) described face characteristic extracts, and specifically may further comprise the steps:
21) to the regular facial image of input,, obtain wavelet sub-bands at different levels according to the same treatment of step 11);
22) with step 21) the anisotropy wavelet package transforms structure that obtains according to step 12) of the wavelet sub-band that obtains adopts the small echo lifting scheme to carry out anisotropy wavelet package transforms (all having identical anisotropy mapped structure with the facial image that guarantees every width of cloth input), obtains corresponding small echo range coefficient; Step 21), 22) be collectively referred to as the anisotropy double-tree complex wavelet package transforms;
23) the small echo range coefficient that obtains through the anisotropy double-tree complex wavelet package transforms is carried out normalization, the range coefficient in each wavelet sub-band is evenly normalized to 0 to 255 from small to large;
24) range coefficient of different wavelet sub-bands is multiplied by the standard deviation (standard deviation of promptly using each wavelet sub-band range coefficient of average face is weighted the wavelet sub-band range coefficient of this facial image) of corresponding each the wavelet sub-band range coefficient that obtains by step 14), obtains using the face characteristic of this facial image correspondence that the small echo range coefficient represents;
3) described classification judgement specifically may further comprise the steps:
31) the positive gray scale facial image of the rule of the every width of cloth in the regular face database carry out step 2) same treatment, obtain the face characteristic of representing by the small echo range coefficient of every width of cloth facial image correspondence, form the standard faces property data base;
32) facial image to be identified carry out step 2) identical processing, obtain the face characteristic of representing by the small echo range coefficient of this facial image correspondence;
The small echo range coefficient face characteristic that the small echo range coefficient face characteristic of the facial image correspondence to be identified that 33) will import is corresponding with every width of cloth facial image in the regular face characteristic database mates one by one, calculates similarity between any two;
34) the small echo range coefficient face characteristic corresponding with the facial image to be identified of input has the result of people's face of maximum similarity as recognition of face in the regular face characteristic database.
Advantage of the present invention
The objective of the invention is to overcome the weak point of prior art, a kind of face identification method of efficient, robust is provided, when reducing computational complexity greatly, guarantee that recognition of face still has very high accuracy.Face identification method of the present invention is mainly based on the anisotropy double-tree complex wavelet package transforms.Remarkable advantage of the present invention is embodied in:
The first, the anisotropy double-tree complex wavelet package transforms has good effect aspect the extraction face characteristic.At first, the anisotropy double-tree complex wavelet package transforms can extract the small echo range coefficient face characteristic of multi-frequency yardstick from facial image; Secondly, anisotropy is decomposed on the basis of the direction that original dual-tree complex wavelet conversion provides, and has introduced more direction, can extract more multidirectional small echo range coefficient face characteristic from facial image, thereby more local grain feature is provided; Moreover the anisotropy double-tree complex wavelet package transforms has TIME SHIFT INVARIANCE, makes the small echo range coefficient have stability preferably on face characteristic extracts.
Second, the small echo range coefficient is weighted, make the face characteristic of representing by the small echo range coefficient that extracts when carrying out recognition of face, have good robustness, can adapt to the facial image that obtains under the different condition, for example different illumination conditions and different expressions.
The 3rd, the computation complexity of this method is very low.At first, the anisotropy mapped structure of average face is directly used in remaining facial image, has significantly reduced the operand of definite anisotropy mapped structure; Secondly, the anisotropy double-tree complex wavelet package transforms can adopt the small echo lifting scheme to realize, make use the anisotropy double-tree complex wavelet package transforms to carry out face characteristic to extract this step operation quick; Moreover the image size of using the small echo range coefficient face characteristic that the anisotropy double-tree complex wavelet package transforms obtains only is the twice of protoplast's face image size, and when classification judgement stage diagram picture mated between any two, calculated amount was very little.
Description of drawings
Fig. 1 is a gray scale facial image pretreatment process block diagram;
Fig. 2 (a) is to the processing of average face and regular facial image feature extraction FB(flow block); Fig. 2 (b) is a recognition of face classification judgement FB(flow block);
Fig. 3 (a) has showed a wavelet sub-band; Fig. 3 (b) expression is carried out the wavelet sub-band of two parts up and down that the anisotropy wavelet package transforms of vertical mapping mode obtains to this wavelet sub-band; Fig. 3 (c) expression is carried out left and right sides two parts wavelet sub-band that the anisotropy wavelet package transforms of horizontal transformation mode obtains to this wavelet sub-band;
Fig. 4 is an embodiment of the present invention, has showed a width of cloth gray scale facial image is carried out the wavelet sub-bands at different levels that 4 grades of dual-tree complex wavelet conversion obtain and the directivity of each wavelet sub-band;
Fig. 5 is an embodiment of the present invention, and wherein, Fig. 5 (a) has showed a width of cloth average face image, and Fig. 5 (b) has showed the mapped structure of anisotropy double-tree complex wavelet package transforms of this average face correspondence and the small echo range coefficient figure that conversion obtains.
Embodiment
The face identification method based on the anisotropy double-tree complex wavelet package transforms that the present invention proposes comprises that processing, the face characteristic to average face extracts and classification judgement three parts, and FB(flow block) reaches embodiment in conjunction with the accompanying drawings and is described in detail as follows as shown in Figure 2:
1) to the processing of average face, specifically may further comprise the steps:
10) the regular positive gray scale facial image of several (being generally more than tens width of cloth) different people in the regular face database is averaged obtain average face;
11) adopt the small echo lifting scheme to carry out J level (J is a positive integer) dual-tree complex wavelet conversion to the average face image and obtain wavelet sub-bands at different levels, the real part of described wavelet sub-band and imaginary part have 2 low frequency wavelet subbands and the 1st, 2 of J+1 level respectively ..., each 6 high frequency wavelet subband of J level, every grade of high frequency wavelet subband all has the wavelet sub-band with 6 kinds of different directions characteristics, and these 6 directions are respectively-75 ° ,-45 °,-15 °, 15 °, 45 °, 75 °;
12) to wherein the 1st, 2, .., J-1 level high frequency wavelet subband adopts the small echo lifting scheme to carry out the anisotropy wavelet package transforms (promptly to j level (j=1 according to variance criterion, 2, .., J-1) the high frequency wavelet subband carries out J-j level anisotropy wavelet package transforms), obtain anisotropy wavelet package transforms structure and corresponding small echo range coefficient; Specifically comprise:
121) the anisotropy wavelet package transforms that the real part part and the imaginary part part of each wavelet sub-band (one of them wavelet sub-band is shown in Fig. 3 (a)) are carried out vertical direction simultaneously, the wavelet sub-band of two parts up and down that calculates (shown in Fig. 3 (b)) range coefficient (being that the real part of wavelet sub-band coefficient and the quadratic sum of imaginary part are opened radical sign) variance sum is designated as Var_H;
122) the anisotropy wavelet package transforms that the real part part and the imaginary part part of each wavelet sub-band are carried out horizontal direction simultaneously, the left and right sides two parts wavelet sub-band that calculates (shown in Fig. 3 (c)) range coefficient variance sum is designated as Var_V;
123) size of comparison Var_H and two values of Var_V if Var_H is bigger, is then finally carried out the anisotropy wavelet package transforms of vertical direction to this wavelet sub-band; If Var_V is bigger, then finally this wavelet sub-band is carried out the anisotropy wavelet package transforms of horizontal direction; Obtain corresponding small echo range coefficient;
124) to the j level (j=1,2 ..., J-1) wavelet sub-band is according to step 121)-123) repeat J-j time, obtain the anisotropy wavelet package transforms structure that the mapping mode of final small echo range coefficient and each wavelet sub-band is combined into;
Above-mentioned steps 11), 12) be collectively referred to as the anisotropy double-tree complex wavelet package transforms;
13) the small echo range coefficient that obtains through the anisotropy double-tree complex wavelet package transforms is carried out normalization, the range coefficient in each wavelet sub-band is evenly normalized to 0 to 255 from small to large;
14) calculate the standard deviation of each wavelet sub-band range coefficient;
2) face characteristic extracts, and specifically may further comprise the steps:
21) to the regular facial image of input,, obtain wavelet sub-bands at different levels according to the same treatment of step 11);
22) with step 21) the anisotropy wavelet package transforms structure that obtains according to step 12) of the wavelet sub-band that obtains adopts the small echo lifting scheme to carry out anisotropy wavelet package transforms (all having identical anisotropy mapped structure with the facial image that guarantees every width of cloth input), obtains corresponding small echo range coefficient; Step 21), 22) be collectively referred to as the anisotropy double-tree complex wavelet package transforms;
23) the small echo range coefficient that obtains through the anisotropy double-tree complex wavelet package transforms is carried out normalization, the range coefficient in each wavelet sub-band is evenly normalized to 0 to 255 from small to large;
24) range coefficient of different wavelet sub-bands is multiplied by the standard deviation (standard deviation of promptly using each wavelet sub-band range coefficient of average face is weighted the wavelet sub-band range coefficient of this facial image) of corresponding each the wavelet sub-band range coefficient that obtains by step 14), obtains using the face characteristic of this facial image correspondence that the small echo range coefficient represents;
3) classification judgement specifically may further comprise the steps:
31) the positive gray scale facial image of the rule of the every width of cloth in the regular face database carry out step 2) same treatment, obtain the face characteristic of representing by the small echo range coefficient of every width of cloth facial image correspondence, form the standard faces property data base;
32) facial image to be identified carry out step 2) identical processing, obtain the face characteristic of representing by the small echo range coefficient of this facial image correspondence;
The small echo range coefficient face characteristic that the small echo range coefficient face characteristic of the facial image correspondence to be identified that 33) will import is corresponding with every width of cloth facial image in the regular face characteristic database mates one by one, calculates similarity between any two;
34) the small echo range coefficient face characteristic corresponding with the facial image to be identified of input has the result of people's face of maximum similarity as recognition of face in the regular face characteristic database.
One embodiment of the present of invention are described in detail as follows in conjunction with Fig. 2: the facial image that present embodiment uses is regular gray scale facial image.
For the gray scale facial image that obtains, at first, facial image is carried out cutting and facial characteristics registration according to the position of human eye.Facial image all is tailored to the image that resolution is 64 * 64 pixel sizes; Simultaneously, the position according to the eyes of people in the facial image is aimed at the eyes position of different facial images substantially.Then, to the registration of the unified size that obtains facial image, carry out histogram equalization, to eliminate condition effect such as uneven illumination as far as possible.
After above-mentioned steps, carry out the embodiments of the invention step again, comprising:
1) to the processing of average face
10) average face is that regular positive gray scale facial image by 121 people in the AR facial image database averages and obtains, shown in Fig. 5 (a);
11) adopt the small echo lifting scheme to carry out 4 grades of dual-tree complex wavelet conversion to the average face image and obtain wavelet sub-bands at different levels, as shown in Figure 4, among the figure, 1,2,3,4,5 respectively the representative obtain the 1st, 2,3,4,5 grades of wavelet sub-bands, wherein, the 5th grade is the low frequency wavelet subband, the 1st, 2,3,4 grades is the high frequency wavelet subband; A, B, C, D, E, F represent the direction of each wavelet sub-band respectively, are respectively 15 °, and 45 °, 75 ° ,-15 ° ,-45 ° ,-75 °;
12) to the 1st, 2, each high frequency wavelet subband of 3 grades adopts the small echo lifting scheme to carry out the anisotropy wavelet package transforms according to variance criterion, and conversion progression is respectively 3,2,1 grade, specifically comprises:
121) to a certain wavelet sub-band shown in Fig. 3 (a), the real part of this wavelet sub-band coefficient part and imaginary part are partly carried out simultaneously the anisotropy wavelet package transforms of vertical direction, the wavelet sub-band of two parts up and down that calculates (shown in Fig. 3 (b)) range coefficient (being that the real part of wavelet sub-band coefficient and the quadratic sum of imaginary part are opened radical sign again) variance sum is designated as Var_H;
122) the imaginary part part of the real part of this wavelet sub-band part is carried out simultaneously the anisotropy wavelet package transforms of horizontal direction, calculate left and right sides two parts wavelet sub-band (shown in Fig. 3 (c)) range coefficient variance sum, be designated as Var_V;
123) size of comparison Var_H and two values of Var_V if Var_H is bigger, is then finally carried out the anisotropy wavelet package transforms of vertical direction to this wavelet sub-band; If Var_V is bigger, then finally this wavelet sub-band is carried out the anisotropy wavelet package transforms of horizontal direction; Obtain corresponding small echo range coefficient;
124) to the 1st, 2,3 grades of high frequency wavelet subbands are according to step 121)-123) repeat 3 respectively, 2,1 time, obtain the anisotropy wavelet package transforms structure that the mapping mode of final small echo range coefficient and each wavelet sub-band is combined into, Fig. 5 (b) has showed average face has been carried out small echo range coefficient and the anisotropy wavelet package transforms structure that the anisotropy double-tree complex wavelet package transforms obtains, wherein, white line among the figure has been showed the mapped structure that the anisotropy double-tree complex wavelet package transforms obtains, different pieces adheres to different wavelet sub-bands separately, and the gray-scale value of each pixel among the figure except that white line is represented the size of small echo range coefficient numerical value;
Step 11), 12) is collectively referred to as the anisotropy double-tree complex wavelet package transforms;
13) the small echo range coefficient that obtains through the anisotropy double-tree complex wavelet package transforms is carried out normalization, the range coefficient in each wavelet sub-band is evenly normalized to 0 to 255 from small to large;
14) calculate the standard deviation of each wavelet sub-band range coefficient;
2) face characteristic extracts
21) to the regular facial image of input, adopt the small echo lifting scheme to carry out 4 grades of dual-tree complex wavelet conversion, obtain wavelet sub-bands at different levels;
22) with step 21) the anisotropy wavelet package transforms structure (shown in Fig. 5 (b)) that obtains according to step 12) of the wavelet sub-band that obtains adopts the small echo lifting scheme to carry out the anisotropy wavelet package transforms, all have identical anisotropy mapped structure with the facial image that guarantees every width of cloth input, obtain corresponding small echo range coefficient; Step 21), 22) be collectively referred to as the anisotropy double-tree complex wavelet package transforms;
23) the small echo range coefficient that obtains through the anisotropy double-tree complex wavelet package transforms is carried out normalization, the range coefficient in each wavelet sub-band is evenly normalized to 0 to 255 from small to large;
24) range coefficient of different wavelet sub-bands is multiplied by the standard deviation (standard deviation of promptly using each wavelet sub-band range coefficient of average face is weighted the wavelet sub-band range coefficient of this facial image) of corresponding each the wavelet sub-band range coefficient that obtains by step 14), obtains using the face characteristic of this facial image correspondence that the small echo range coefficient represents;
3) classification judgement
31) the positive gray scale facial image of the rule of the every width of cloth in the regular face database carry out step 2) same treatment, obtain the face characteristic of representing by the small echo range coefficient of every width of cloth facial image correspondence, form the standard faces property data base;
32) facial image to be identified carry out step 2) identical processing, obtain the face characteristic of representing by the small echo range coefficient of this facial image correspondence;
The small echo range coefficient face characteristic that the small echo range coefficient face characteristic of the facial image correspondence to be identified that 33) will import is corresponding with every width of cloth facial image in the regular face characteristic database mates one by one, calculate similarity between any two, what adopt here is Euclidean distance (but be not limited only to Euclidean distance, also can adopt a norm apart from the tolerance mode of tolerance similarity that waits other);
34) the small echo range coefficient face characteristic corresponding with the facial image to be identified of input has the result of people's face of maximum similarity as recognition of face in the regular face characteristic database.
But the present invention is not restricted to the described concrete grammar of the foregoing description, and all any conversion of carrying out according to technology contents of the present invention to present embodiment all should belong to protection category of the present invention.
Experimental result for example
Adopt AR facial image database (referring to A.M.Martinez, and R.Benavente.The AR Face Database.Technique Report#24CVC., 1998.), this database comprises 121 people, everyone has the positive gray scale facial image of width of cloth rule, composition rule facial image database; Everyone also respectively has the gray scale facial image of two width of cloth expression shape change and the gray scale facial image of three width of cloth illumination variation, and these facial images are used to test the recognition correct rate of different people face recognition method as facial image to be identified;
1) for 242 width of cloth facial image of expression shape change, adopting the recognition correct rate of existing Gabor conversion face identification method is 96.28%, and the recognition correct rate based on the face identification method of anisotropy double-tree complex wavelet package transforms that adopts that the present invention proposes is 96.69%;
2) for 363 width of cloth facial images of illumination variation, adopting the recognition correct rate of existing Gabor conversion face identification method is 97.52%, and the recognition correct rate based on the face identification method of anisotropy double-tree complex wavelet package transforms that adopts that the present invention proposes is 97.52%;
3) face characteristic of extraction one width of cloth facial image is respectively employed averaging time: 32 milliseconds consuming time of existing Gabor conversion face identification method, 18 milliseconds consuming time of the face identification method based on the anisotropy double-tree complex wavelet package transforms that the present invention proposes;
4) two width of cloth face characteristics are mated be respectively employed averaging time: 0.33 millisecond consuming time of existing Gabor conversion face identification method, 0.02 millisecond consuming time of the face identification method that the present invention proposes based on the anisotropy double-tree complex wavelet package transforms.
Claims (2)
1, a kind of face identification method based on the anisotropy double-tree complex wavelet package transforms is characterized in that, comprises that processing, the face characteristic to average face extracts and classification judgement three parts:
1) described processing to average face specifically may further comprise the steps:
10) the regular positive gray scale facial image of several different people in the regular face database is averaged obtain average face;
11) adopt the small echo lifting scheme to carry out the conversion of J level dual-tree complex wavelet to the average face image and obtain wavelet sub-bands at different levels, J is a positive integer, and described wavelet sub-band has 2 low frequency wavelet subbands and the 1st, 2 of J+1 level ..., each 6 high frequency wavelet subband of J level;
12) to wherein the 1st, 2 ..., J-1 level high frequency wavelet subband adopts the small echo lifting scheme to carry out the anisotropy wavelet package transforms according to variance criterion, obtains anisotropy wavelet package transforms structure and corresponding small echo range coefficient;
Above-mentioned steps 11), 12) be collectively referred to as the anisotropy double-tree complex wavelet package transforms;
13) the small echo range coefficient that obtains through the anisotropy double-tree complex wavelet package transforms is carried out normalization, the range coefficient in each wavelet sub-band evenly is normalized to 0 to 255 from small to large;
14) calculate the standard deviation of each wavelet sub-band range coefficient;
2) described face characteristic extracts, and specifically may further comprise the steps:
21) to the regular facial image of input,, obtain wavelet sub-bands at different levels according to the same treatment of step 11);
22) with step 21) the anisotropy wavelet package transforms structure that obtains according to step 12) of the wavelet sub-band that obtains adopts the small echo lifting scheme to carry out the anisotropy wavelet package transforms, obtains corresponding small echo range coefficient; Step 21), 22) be collectively referred to as the anisotropy double-tree complex wavelet package transforms;
23) the small echo range coefficient that obtains through the anisotropy double-tree complex wavelet package transforms is carried out normalization, the range coefficient in each wavelet sub-band is evenly normalized to 0 to 255 from small to large;
24) range coefficient of different wavelet sub-bands is multiplied by the standard deviation of corresponding each the wavelet sub-band range coefficient that obtains by step 14), obtains using the face characteristic of this facial image correspondence that the small echo range coefficient represents;
3) described classification judgement specifically may further comprise the steps:
31) the positive gray scale facial image of the rule of the every width of cloth in the regular face database carry out step 2) same treatment, obtain the face characteristic of representing by the small echo range coefficient of every width of cloth facial image correspondence, form the standard faces property data base;
32) facial image to be identified carry out step 2) identical processing, obtain the face characteristic of representing by the small echo range coefficient of this facial image correspondence;
The small echo range coefficient face characteristic that the small echo range coefficient face characteristic of the facial image correspondence to be identified that 33) will import is corresponding with every width of cloth facial image in the regular face characteristic database mates one by one, calculates similarity between any two;
34) the small echo range coefficient face characteristic corresponding with the facial image to be identified of input has the result of people's face of maximum similarity as recognition of face in the regular face characteristic database.
2, the method for claim 1 is characterized in that, described step 12) obtains anisotropy wavelet package transforms structure and corresponding small echo range coefficient, specifically comprises:
Described wavelet sub-band at different levels has real part and imaginary part, the real part of described wavelet sub-band and imaginary part have 2 low frequency wavelet subbands and the 1st, 2 of J+1 level respectively ..., each 6 high frequency wavelet subband of J level, every grade of high frequency wavelet subband all has the wavelet sub-band with 6 kinds of different directions characteristics, and these 6 directions are respectively-75 ° ,-45 °,-15 °, 15 °, 45 °, 75 °;
121) the anisotropy wavelet package transforms that the real part part and the imaginary part part of each wavelet sub-band are carried out vertical direction simultaneously, the wavelet sub-band of two parts up and down range coefficient that calculates (being that the real part of wavelet sub-band coefficient and the quadratic sum of imaginary part are opened radical sign) variance sum is designated as Var_H;
122) the anisotropy wavelet package transforms that the real part part and the imaginary part part of each wavelet sub-band are carried out horizontal direction simultaneously, the left and right sides two parts wavelet sub-band range coefficient variance sum that calculates is designated as Var_V;
123) size of comparison Var_H and two values of Var_V if Var_H is bigger, is then finally carried out the anisotropy wavelet package transforms of vertical direction to this wavelet sub-band; If Var_V is bigger, then finally this wavelet sub-band is carried out the anisotropy wavelet package transforms of horizontal direction; Obtain corresponding small echo range coefficient;
124) to the j level (j=1,2 ..., J-1) wavelet sub-band is according to step 121)-123) repeat J-j time, obtain the anisotropy wavelet package transforms structure that the mapping mode of final small echo range coefficient and each wavelet sub-band is combined into.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CNB2008101063962A CN100557623C (en) | 2008-05-13 | 2008-05-13 | Face identification method based on the anisotropy double-tree complex wavelet package transforms |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CNB2008101063962A CN100557623C (en) | 2008-05-13 | 2008-05-13 | Face identification method based on the anisotropy double-tree complex wavelet package transforms |
Publications (2)
Publication Number | Publication Date |
---|---|
CN101271521A true CN101271521A (en) | 2008-09-24 |
CN100557623C CN100557623C (en) | 2009-11-04 |
Family
ID=40005483
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CNB2008101063962A Expired - Fee Related CN100557623C (en) | 2008-05-13 | 2008-05-13 | Face identification method based on the anisotropy double-tree complex wavelet package transforms |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN100557623C (en) |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102411708A (en) * | 2011-12-02 | 2012-04-11 | 湖南大学 | Face recognition method combining dual-tree complex wavelet transform and discrete wavelet transform |
CN102509138A (en) * | 2011-11-18 | 2012-06-20 | 山东神思电子技术股份有限公司 | Authentication system based on second-generation ID card and human face feature recognition and working method therefor |
WO2015090126A1 (en) * | 2013-12-16 | 2015-06-25 | 北京天诚盛业科技有限公司 | Facial characteristic extraction and authentication method and device |
WO2015154205A1 (en) * | 2014-04-11 | 2015-10-15 | Xiaoou Tang | Methods and systems for verifying face images based on canonical images |
CN106092879A (en) * | 2016-06-07 | 2016-11-09 | 西安向阳航天材料股份有限公司 | Explosion clad pipe bonding state detection method based on vibratory response information |
WO2017114133A1 (en) * | 2015-12-31 | 2017-07-06 | 深圳光启合众科技有限公司 | Image recognition method and device |
WO2017181952A1 (en) * | 2016-04-19 | 2017-10-26 | 中兴通讯股份有限公司 | Positioning method and device |
CN109117880A (en) * | 2018-08-06 | 2019-01-01 | 广东工业大学 | Tile image sorting algorithm based on WAVELET PACKET DECOMPOSITION selection coefficient weighting reconstruct |
CN109740423A (en) * | 2018-11-22 | 2019-05-10 | 霍尔果斯奇妙软件科技有限公司 | Ethnic recognition methods and system based on face and wavelet packet analysis |
CN109829381A (en) * | 2018-12-28 | 2019-05-31 | 北京旷视科技有限公司 | A kind of dog only identifies management method, device, system and storage medium |
-
2008
- 2008-05-13 CN CNB2008101063962A patent/CN100557623C/en not_active Expired - Fee Related
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102509138A (en) * | 2011-11-18 | 2012-06-20 | 山东神思电子技术股份有限公司 | Authentication system based on second-generation ID card and human face feature recognition and working method therefor |
CN102509138B (en) * | 2011-11-18 | 2014-07-02 | 山东神思电子技术股份有限公司 | Authentication system based on second-generation ID card and human face feature recognition and working method therefor |
CN102411708A (en) * | 2011-12-02 | 2012-04-11 | 湖南大学 | Face recognition method combining dual-tree complex wavelet transform and discrete wavelet transform |
WO2015090126A1 (en) * | 2013-12-16 | 2015-06-25 | 北京天诚盛业科技有限公司 | Facial characteristic extraction and authentication method and device |
WO2015154205A1 (en) * | 2014-04-11 | 2015-10-15 | Xiaoou Tang | Methods and systems for verifying face images based on canonical images |
CN106462724B (en) * | 2014-04-11 | 2019-08-02 | 北京市商汤科技开发有限公司 | Method and system based on normalized images verification face-image |
CN106462724A (en) * | 2014-04-11 | 2017-02-22 | 北京市商汤科技开发有限公司 | Methods and systems for verifying face images based on canonical images |
WO2017114133A1 (en) * | 2015-12-31 | 2017-07-06 | 深圳光启合众科技有限公司 | Image recognition method and device |
WO2017181952A1 (en) * | 2016-04-19 | 2017-10-26 | 中兴通讯股份有限公司 | Positioning method and device |
CN107305246A (en) * | 2016-04-19 | 2017-10-31 | 中兴通讯股份有限公司 | Localization method and device based on received signal strength indicator |
CN107305246B (en) * | 2016-04-19 | 2023-08-08 | 中兴通讯股份有限公司 | Positioning method and device based on received signal strength indication |
CN106092879B (en) * | 2016-06-07 | 2019-07-12 | 西安向阳航天材料股份有限公司 | Explosion clad pipe bonding state detection method based on vibratory response information |
CN106092879A (en) * | 2016-06-07 | 2016-11-09 | 西安向阳航天材料股份有限公司 | Explosion clad pipe bonding state detection method based on vibratory response information |
CN109117880A (en) * | 2018-08-06 | 2019-01-01 | 广东工业大学 | Tile image sorting algorithm based on WAVELET PACKET DECOMPOSITION selection coefficient weighting reconstruct |
CN109740423A (en) * | 2018-11-22 | 2019-05-10 | 霍尔果斯奇妙软件科技有限公司 | Ethnic recognition methods and system based on face and wavelet packet analysis |
CN109829381A (en) * | 2018-12-28 | 2019-05-31 | 北京旷视科技有限公司 | A kind of dog only identifies management method, device, system and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN100557623C (en) | 2009-11-04 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN100557623C (en) | Face identification method based on the anisotropy double-tree complex wavelet package transforms | |
Cao et al. | Illumination invariant extraction for face recognition using neighboring wavelet coefficients | |
CN101329724B (en) | Optimized human face recognition method and apparatus | |
CN101866421B (en) | Method for extracting characteristic of natural image based on dispersion-constrained non-negative sparse coding | |
Berbar | Three robust features extraction approaches for facial gender classification | |
Bouchaffra et al. | Structural hidden Markov models for biometrics: Fusion of face and fingerprint | |
Rai et al. | A gender classification system robust to occlusion using Gabor features based (2D) 2PCA | |
CN101241550B (en) | Iris image quality judgment method | |
CN102222215A (en) | Gait recognition method based on two-dimension wavelet packet decomposition and complete PCA (Principal Component Analysis) | |
Goh et al. | Wavelet based illumination invariant preprocessing in face recognition | |
CN103942545A (en) | Method and device for identifying faces based on bidirectional compressed data space dimension reduction | |
Emadi et al. | Illumination normalization using 2D wavelet | |
Zhao et al. | Face recognition feature comparison based SVD and FFT | |
Viswanathan et al. | Fingerprint enhancement and compression method using Morlet wavelet | |
Meraoumia et al. | Do multispectral palmprint images be reliable for person identification? | |
Priya et al. | Local fusion of complex dual-tree wavelet coefficients based face recognition for single sample problem | |
Lokhande et al. | Wavelet packet based iris texture analysis for person authentication | |
Islam et al. | Partial iris image recognition using wavelet based texture features | |
Patil et al. | Performance improvement of face recognition system by decomposition of local features using discrete wavelet transforms | |
CN109711305A (en) | Merge the face identification method of a variety of component characterizations | |
Zhang et al. | Study of finite radon transform in face recognition | |
CN102831406A (en) | Human face identification method based on spatial filters | |
Fan et al. | Improved wavelet-based illumination normalization algorithm for face Recognition | |
Xuebin et al. | NSCTWavelet: An efficient method for multimodal biometric recognition based on pixel level fusion | |
Haritha et al. | Face Recognition System Using Doubly Truncated Multivariate Gaussian Mixture Model and DCT Coefficients Under Logarithm Domain |
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 | ||
C17 | Cessation of patent right | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20091104 Termination date: 20110513 |