CN104463137A - Anomaly facial image detection method and system based on characteristic space decomposition - Google Patents

Abstract

Disclosed is an anomaly facial image detection method and system based on characteristic space decomposition. Principal component analysis is conducted on data in a face database to obtain a characteristic space model containing the principle direction, then the leave-one-out method is adopted to take each image in a data set as a candidate sample to be deleted from the data set in sequence, the characteristic space model is updated by means of characteristic space decomposition, angle-based anomaly detection is conducted on multiple principle directions of old and new characteristic spaces, and finally anomaly factors are calculated through the diagonal element of a rotation matrix during characteristic space decomposition and an anomaly facial image is detected.

Description

The abnormal face image detecting method of feature based space decomposition and system

Technical field

What the present invention relates to is a kind of technology of image processing field, specifically a kind of can improve the detection perform of prior art on high dimension and large scale data and counting yield, the detection method of the abnormal face image of feature based space decomposition and system.

Background technology

Along with the progress of infotech and imaging sensor aspect, and computing power promotes, increasing facial image or face video appear in people's life, such as, based on the gate control system of recognition of face, the supervisory system etc. based on face video tracking.This type systematic is when implementing, and in order to obtain satisfied performance, the facial image database that usual Water demand one is very large, trains and extract faceform; The recognition accuracy of faceform's guarantee system accurately or tracking effect.But, in actual life, in the building process of training facial image database, due to the change of image imaging circumstances, sensor noise and human factor etc., facial image database is often mixed into some containing excessive noise or by the underproof image seriously blocked, or even non-face image.Many faceforms are very sensitive to abnormal data, and that is a small amount of abnormal face image may have a strong impact on the accuracy of final faceform.Usually, these images can manually be deleted when making database.But along with the development of large data and social networks, the scale of facial image database is increasing, and such mode of operation needs the human cost of at substantial.Thus, in the urgent need to a kind of can machine realize Outlier Detection Algorithm, it can find to be mixed in those abnormal images in a facial image database.This just this patent solve technical matters.

Abnormal data is defined as by Hawkins " observation departs from other and observes undue far away, to such an extent as to it is generated by other mechanism to allow people suspect ".Large data age data present the feature of high dimension and large scale (large sample amount) usually.Traditional Outlier Detection Algorithm needs tentation data to obey some specific distributions usually, or need to generate each data sample k ?arest neighbors, thus can run into so-called " dimension crisis " problem.In high-dimensional environment, the observation dimension d of data will be far longer than the number n of viewed sample usually.Thus, when dimension is larger, less data sample seems very sparse in data space, to such an extent as to range information becomes unreliable, and the concept of neighborhood also becomes nonsensical.When lacking prior imformation, from data, find that its distributed model is very difficult.In addition, in higher-dimension or large scale problem, generate each data k ?the calculating of arest neighbors be very consuming time.Abnormality detection in facial image database is a typical higher-dimension and large scale problem, and namely the dimension of data is very high, and the number of sample is very large.Thus, design a kind of abnormal face image detection algorithm that is accurate, quick, robust and remain a challenging task and urgent need to solve the problem.

Through finding the retrieval of prior art, Barnett [Barnett V.The ordering of multivariate data (withdiscussion) [J] .Journal of the royalstatistical society, 1976, Series.A (139): 318 – 354] propose a kind of T based on Hotelling ²statistics quantitative statistics method for detecting abnormality, its key step is: 1) think that data entirety obeys a multivariate normal distribution, then use training data to calculate its average and variance matrix; 2) utilize the abnormal score of mahalanobis distance calculated candidate sample, it obeys the T of Hotelling ²distribution, if score is excessive, this candidate samples is an exceptional sample.The major defect of the method is, if the dimension of data is higher, its applicability is limited, because be difficult to determine rational Data distribution8.Breunig etc. propose a kind of method for detecting abnormality of density based, i.e. local outlier factor LOF [Breunig M, Kriegel H P, Ng R T, et al.LOF:Identifying density ?based local outliers.Proceeding of the 2000ACM SIGMOD InternationalConference on Management ofData ACM, 2000:93 – 104], it determines its abnormality degree according to the local density of each sample, thus can excavate and be hidden in local anomaly in data global structure, its key step is, first calculate each sample k ?neighborhood, calculate local density in neighborhood again, if the local density of a point is too large, so that is exactly an abnormity point.The major defect of LOF is that counting yield is lower on higher-dimension or large scale data because need to calculate each sample k ?neighborhood.The people such as Kriegel propose a kind of method for detecting abnormality based on angle, be called angle Outlier factor fastABOF [Kriegel H P, hubert M S, Zimek A.Angle ?based outlier detection in high ?dimensionaldata.Proceeding of the14th ACM SIGKDD international conference on Knowledgediscovery and data mining ACM, 2008:444 – 452], its key step is, first calculate each sample k ?neighborhood, calculate the angle that in neighborhood, any two points and this point are formed again, then calculate the variance of angled cosine value, if the variance that sample is corresponding is very little, so it is an exceptional sample.FastABOD main determines it is that its counting yield in large scale data is lower because it need to calculate each sample k ?neighborhood.

Open (bulletin) the day 2013.10.23 of Chinese patent literature CN103365969A, disclose a kind of method and system of anomaly data detection process, its key step is: gather the particular data in the schedule time, extract the characteristic information in particular data again, determine the data space of data, if the data of a last sample have exceeded corresponding data space, so it has been abnormal data.The major defect of the method is, when the dimension of data is very high, only judges that the Detection results of abnormality is limited according to the value region of every one dimension, in addition the selective dependency experience of the characteristic area of every one dimension.

Summary of the invention

The present invention is directed on higher-dimension and large scale problem that existing method detects at abnormal face not enough, a kind of abnormal face image detecting method of feature based space decomposition is proposed, can detect containing excessive noise from given facial image database and be verified and block abnormal face image, or non-face image.On the higher-dimension large scale data set that facial image database is such, it has higher susceptibility to an exceptional sample, thus can improve detection perform and the counting yield of abnormal face image.Basic thought of the present invention utilizes principal component analysis (PCA) to the susceptibility of abnormal data, under leaving-one method mode, using the criterion of the change size of the principal direction caused by deletion candidate samples as abnormality detection.In order to accurately calculate the knots modification of the principal direction that a candidate face image causes, have employed feature space division with the principal component analysis (PCA) of the more new data set of decrement fast, and its result is optimum in theory.The present invention also finds that the diagonal entry of the rotation matrix of the principal direction in feature space splitting-up method equals the cosine similarity of principal direction angle knots modification.Thus, the present invention detects eventually through observing the diagonal entry value of rotation matrix when feature space divides the abnormal face image be hidden in face database.

The abnormal face image detecting method of the feature based space decomposition that the present invention proposes comprises the steps:

Step one: claim row by the two-dimensional data matrix of the every piece image in facial image database, obtain a vector; Then facial image database can be designated as a data set X={x containing n observation sample ₁, x ₂..., x _n, wherein x _i∈ R ^d;

Step 2: principal component analysis (PCA) is calculated to data centralization total data, obtains its feature space model;

Step 3: adopt leaving-one method, select each sample alternatively sample successively, then by it from data set from deletion;

Step 4: the feature space model adopting feature space splitting-up method more new data set, record rotation matrix wherein;

Step 5: the diagonal entry extracting rotation matrix, then uses the detection method based on angle, judge whether this candidate samples is abnormal face image.

Feature space model in described step 2 refers to: the data set X={x containing n observation sample for ₁, x ₂..., x _n, wherein x _i∈ R ^dthe feature space model of X is the feature decomposition of its covariance matrix, be denoted as Ω (X)=(N (X), μ (X), Λ (X), U (X)), wherein: N (X) is the number of observation sample in X, μ (X) is the average of sample, and Λ (X) is the individual maximum eigenwert { λ of front k of X Eigenvalues Decomposition ₁, λ ₂..., λ _kthe diagonal matrix that forms, U (X) is these eigenwert characteristics of correspondence vectors, i.e. principal direction.

Feature space division in described step 4 refers to: use leaving-one method to detect exceptional sample, at every turn by a candidate samples x _tdelete from complete data centralization, then abnormality detection is carried out in the change by observing principal direction in feature space, remaining data after deleting candidate samples integrates as Z, its characteristic of correspondence spatial model is Ω (Z)=(N (Z), μ (Z), Λ (Z), U (Z)), in higher-dimension large scale situation, the calculated amount using the feature decomposition of the covariance matrix of Z to calculate Ω (Z) is larger, present invention employs a kind of update method of decrement accordingly, by carrying out to Ω (X) the quick and precisely calculating that little amendment achieves Ω (Z): note candidate samples integrates as Y={x _t, obviously not needing to calculate the feature space model just obtaining it is Ω (Y)=(1, x _t, 0,0), so, Ω (Z) can be regarded as and divide the later result of Ω (Y) from Ω (X), thus can be expressed as

Described division process, specifically comprises the following steps:

N(Z)＝n-1

$\mathrm{\μ}\left(Z\right)=\frac{\mathrm{n\μ}\left(X\right)-x_t}{n-1}$

$\widehat{\mathrm{\Ω}}=\frac{n}{n-1}\mathrm{\Λ}\left(X\right)-\frac{1}{n-1}{\mathrm{gg}}^T\stackrel{\mathrm{EVD}}{\⇒}\mathrm{R\Λ}\left(X\right)R^T$

U(Z)＝U(X)R

Wherein: g=U (X) ^t(X _t-μ (X)), EVD representation feature value is decomposed.

The abnormal face image based on angle in described step 5 detects and refers to: by set up feature space division and based on angle abnormality detection between natural link, whether the diagonal entry by means of only inspection rotation matrix can candidate samples be just abnormal data, and calculated amount is very little: when the principal direction of the feature space model Ω (X) of former data set is U (X)={ u ₁, u ₂..., u _k, delete candidate samples x _tthe principal direction of the feature space model Ω (Z) of rear remaining data be U (Z)=u ' ₁, u ' ₂..., u ' _k.Thus, the cosine similarity of corresponding principal direction is according to feature space splitting-up method, U (Z) be U (X) by rotational transform R gained, u ' m=U (X) r _m, wherein r _mthe m row of R.So, u _m ^tu ' _m=R _mm, wherein R _mmm the diagonal entry of rotation matrix R.That is, the diagonal entry of rotation matrix R is exactly the cosine similarity that feature space model principal direction changes.Finally, candidate samples x _toutlier factor be: if S _tcomparatively large, candidate samples x is described _tcause the change of larger principal direction, so x _tit is exactly an exceptional sample.

Technique effect

Compared with prior art, technique effect of the present invention mainly: the first, divided by feature space, the present invention do not need to calculate each sample k ?neighborhood, thus counting yield is very high, especially on higher-dimension and large scale data set; The second, based on the angle method for detecting abnormality of many principal direction strategy, in the feature space of low-dimensional, use angle information considers abnormality, makes the detection perform of the present invention on high dimensional data obviously due to prior art.

Accompanying drawing explanation

Fig. 1 is schematic diagram of the present invention.

Fig. 2 is many major components strategy schematic diagram.

Fig. 3 is embodiment particular flow sheet.

Embodiment

Elaborate to embodiments of the invention below, the present embodiment is implemented under premised on technical solution of the present invention, give detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

Embodiment 1

As shown in Figure 1, be the implementation method of the present embodiment, the present embodiment comprises the following steps:

Step one: claim row by the two-dimensional data matrix of the every piece image in facial image database, obtain a vector; Then facial image database can be designated as a data set X={x containing n observation sample ₁, x ₂..., x _n, wherein x _i∈ R ^d;

Step 2: principal component analysis (PCA) is calculated to data centralization total data, obtains its feature space model;

Step 3: adopt leaving-one method, select each sample alternatively sample successively, then by it from data set from deletion;

Step 4: the feature space model adopting feature space splitting-up method more new data set, record rotation matrix wherein;

Step 5: the diagonal entry extracting rotation matrix, then uses the detection method based on angle, judge whether this candidate samples is abnormal face image.

As shown in Figure 3, in order to detect data-oriented collection X={x ₁, x ₂..., x _nin exceptional data point, the concrete implementation step of the present embodiment is as follows:

1) its feature space Ω (X)=(N (X), μ (X), Λ (X), U (X)) is obtained by the Eigenvalues Decomposition of the covariance matrix calculating X.

2) to each sample x in X _t, t=1,2 ..., n.Use feature space division decrement regeneration characteristics subspace

3) average upgrades μ (Z)=(n μ (X)-x _t)/(n-1);

4) the projection g=U (X) of computation of mean values residual error ^t(x _t-μ (X));

5) covariance matrix on feature space is calculated

6) calculate eigenvalues Decomposition R Λ (X) R ^t.

7) use many principal direction strategy, the angle changed by principal direction calculates Outlier factor:

Described multiple principal directions, namely many principal direction strategy refers to: when data centralization sample number is very large, and in order to improve the detection perform of algorithm on big-sample data collection, the knots modification of first principal direction may be not enough to the existence observing abnormity point.Multiple principal direction strategy deletes the knots modification of front l principal direction caused by a candidate samples under considering leaving-one method pattern.The basic thought of this strategy is that secondary leading principal direction is more responsive to abnormity point than first principal direction.

As shown in Figure 2, for many principal direction strategy is used for the Visual Explanation of abnormality detection.In the left side subgraph of Fig. 2 (a), three solid lines represent first three principal direction of the feature space model of raw data set respectively.In the right subgraph of Fig. 2 (a), right side round dot represents a normal sample, now it is deleted from data centralization, and the small change of the direction generation of major component, has rotated to new solid line position from original principal direction from dotted line position.In the right side subgraph of Fig. 2 (b), round dot represents an exceptional sample, and it deviate from the region of normal sample.Now it deleted from data centralization, so can there is larger change in the principal direction of feature space model, wherein: the anglec of rotation of first principal direction is less than the anglec of rotation of time principal direction.

8) cosine similarity of the corresponding principal direction of new and old feature space equals the diagonal entry R of rotation matrix R _mm;

9) sample x is calculated _toutlier factor if S _tbe greater than threshold value T _d, note x _tfor exceptional sample, otherwise be designated as normal sample.

The present embodiment relates to a kind of system realizing said method, comprise: principal component analysis (PCA) module, feature space division module and angle anomaly data detection module, wherein: principal component analysis (PCA) module calculates the feature space model of full dataset and exports feature space division module to and carry out model modification, and angle anomaly data detection module is carried out angle according to feature space model after the feature space from principal component analysis (PCA) module and the renewal from feature space division module and detected also output abnormality Data Detection result.

Described feature space division module comprises: average updating block, average Residual projection unit, feature space covariance matrix unit and Eigenvalues Decomposition unit, wherein: average updating block is connected with average Residual projection unit and transmits the residual error of deleting caused by candidate samples, average Residual projection unit is connected with feature space covariance matrix unit and transmits covariance information, feature space covariance matrix unit is connected with Eigenvalues Decomposition unit and transmits covariance matrix, Eigenvalues Decomposition unit is connected with angle anomaly data detection module and transmits rotation matrix information.

Described angle anomaly data detection module comprises: diagonal entry extraction unit, Outlier factor computing unit and threshold value comparing unit, wherein: diagonal entry extraction unit is connected with Outlier factor computing unit and transmits the cosine similarity information of new and old principal direction, Outlier factor computing unit is connected with threshold value comparison model and transmission abnormality factor information, Outlier factor and threshold value compare by threshold value comparing unit, output detections result, namely this candidate samples is normal data or abnormal data.

The testing result of the anomaly data detection algorithm of the feature based space decomposition that table 1, table 2 and table 3 propose for the present embodiment (ES ?OD) in two groups of real human face image libraries commonly used.Table 1 is the specifying information of adopted test data.In order to embody the detection perform of this algorithm, test to ES ?OD and traditional two kind of Outlier Detection Algorithm LOF and FastABOD compare.In order to comprehensively weigh the performance of several Outlier Detection Algorithm, adopt AUC score and computing time as evaluation index, result is respectively in table 2 and table 3.No matter can find out, be that detection perform is crossed or counting yield, the ES of the present embodiment ?OD be all higher than two kinds of classic methods.

Table 1 test data configures

	Dimension	Sample number
			FERET	644	720
AR	644	952

The ES of table 2 LOF, FastABOD and the present embodiment ?the abnormality detection AUC score of OD

	LOF	FastABOD	ES‐OD
				FERET	0.9398	0.8861	0.9248
AR	0.9715	0.9679	0.9869

The ES of table 3 LOF, FastABOD and the present embodiment ?abnormality detection computing time of OD

	LOF	FastABOD	ES‐OD
				FERET	7.4672	15.812	1.7460
AR	15.432	25.959	2.2726

Claims (8)

1. an abnormal face image detecting method for feature based space decomposition, is characterized in that, by carrying out to the data in a face database feature space model that principal component analysis (PCA) obtains comprising principal direction; Then adopt leaving-one method, successively by each image alternatively sample of data set, it is deleted from data centralization; Feature space is utilized to divide regeneration characteristics spatial model; Abnormality detection based on angle is carried out to multiple principal directions of new and old feature space, finally extracts abnormal face image.

2. method according to claim 1, is characterized in that, described feature space model refers to: the data set X={x containing n observation sample for ₁, x ₂..., x _n, wherein x _i∈ R ^dthe feature space model of X is the feature decomposition of its covariance matrix, be denoted as Ω (X)=(N (X), μ (X), Λ (X), U(X)), wherein: N (X) is the number of observation sample in X, μ (X) is the average of sample, and Λ (X) is the individual maximum eigenwert { λ of front k of X Eigenvalues Decomposition ₁, λ ₂..., λ _kthe diagonal matrix that forms, U (X) is these eigenwert characteristics of correspondence vectors, i.e. principal direction.

3. method according to claim 1, is characterized in that, described feature space division refers to: use leaving-one method to detect exceptional sample, at every turn by a candidate samples x _tdelete from complete data centralization, then abnormality detection is carried out in the change by observing principal direction in feature space, remaining data after deleting candidate samples integrates as Z, its characteristic of correspondence spatial model is Ω (Z)=(N (Z), μ (Z), Λ (Z), U (Z)), its feature space model is Ω (Y)=(1, x _t, 0,0), be expressed as further wherein: Y={x _tit is candidate samples collection.

4. method according to claim 1, is characterized in that, described division process, specifically comprises:

N(Z)＝n-1；

$\mathrm{\μ}\left(Z\right)=\frac{\mathrm{n\μ}\left(X\right)-x_t}{n-1};$

$\widehat{\mathrm{\Ω}}=\frac{n}{n-1}\mathrm{\Λ}\left(X\right)-\frac{1}{n-1}{\mathrm{gg}}^T\stackrel{\mathrm{EVD}}{\⇒}\mathrm{R\Λ}\left(X\right)R^T;$

U(Z)＝U(X)R；

Wherein: g=U (X) ^t(x _t-μ (X)), EVD representation feature value is decomposed.

5. method according to claim 1, it is characterized in that, the described abnormality detection based on angle refers to: by set up feature space division and based on angle abnormality detection between natural link, whether the diagonal entry by means of only inspection rotation matrix can candidate samples be just abnormal data, and concrete steps comprise:

1) when the principal direction of the feature space model Ω (X) of former data set is U (X)={ u ₁, u ₂..., u _k, delete candidate samples x _tthe principal direction of the feature space model Ω (Z) of rear remaining data be U (Z)=u ' ₁, u ' ₂..., u ' _k; The cosine similarity of corresponding principal direction is $s_m=1-\frac{<u_m,u_m^{\&prime;}>}{\left|\right|u_m\left|\right|\left|\right|u_m^{\&prime;}\left|\right|}=1-\left|{u_m}^T{u}_m^{\&prime;}\right|;$

2) according to feature space splitting-up method, U (Z) be U (X) by rotational transform R gained, u ' _m=U (X) r _m, wherein r _mthe m row of R, so u _mtu ' _m=R _mm, wherein R _mmbe m the diagonal entry of rotation matrix R, namely the diagonal entry of rotation matrix R is exactly the cosine similarity that feature space model principal direction changes;

3) candidate samples x _toutlier factor be: work as S _tbe greater than threshold value T _d, then x _tfor exceptional sample.

6. one kind realizes the system of above-mentioned arbitrary claimed method, it is characterized in that, comprise: principal component analysis (PCA) module, feature space division module and angle anomaly data detection module, wherein: principal component analysis (PCA) module calculates the feature space model of full dataset and exports feature space division module to and carry out model modification, and angle anomaly data detection module is carried out angle according to feature space model after the feature space from principal component analysis (PCA) module and the renewal from feature space division module and detected also output abnormality Data Detection result.

7. system according to claim 6, it is characterized in that, described feature space division module comprises: average updating block, average Residual projection unit, feature space covariance matrix unit and Eigenvalues Decomposition unit, wherein: average updating block is connected with average Residual projection unit and transmits the residual error of deleting caused by candidate samples, average Residual projection unit is connected with feature space covariance matrix unit and transmits covariance information, feature space covariance matrix unit is connected with Eigenvalues Decomposition unit and transmits covariance matrix, Eigenvalues Decomposition unit is connected with angle anomaly data detection module and transmits rotation matrix information.

8. system according to claim 6, it is characterized in that, described angle anomaly data detection module comprises: diagonal entry extraction unit, Outlier factor computing unit and threshold value comparing unit, wherein: diagonal entry extraction unit is connected with Outlier factor computing unit and transmits the cosine similarity information of new and old principal direction, Outlier factor computing unit is connected with threshold value comparison model and transmission abnormality factor information, Outlier factor and threshold value compare by threshold value comparing unit, output detections result, namely this candidate samples is normal data or abnormal data.