CN102750546B - Face shielding detection method based on structured error code - Google Patents

Abstract

The invention provides a face shielding detection method based on a structured error code, which has high detection accuracy and good feasibility, and is suitable for dealing with situations of lower image dimension and larger shielding area. The method comprises the following concrete steps of: step 1, stretching detected face image data and training sample data into a column vector; step 2, defining an error support, and initializing; step 3, under a minimum CD (Compact Disc) error standard, calculating a sparse code and a reconstruction error of the detected face image data on a dictionary formed by the training sample data through the error support; step 4, estimating an error support according to the reconstruction error; step 5, building an aspect graph for describing the error support, and estimating an error support again through the aspect graph and the reconfiguration error; step 6, iterating the step 3-the step 5 to obtain a reconfiguration error sequence and an error support sequence; and step 7, selecting the optimal error support, and obtaining a set of occluded pixel points in the detected face image according to the optimal error support.

Description

Face occlusion detection method based on structuring error coding

Technical field

The invention belongs to image processing field, particularly relate to recognition of face field.

Background technology

Along with the high speed development of infotech, face recognition technology has been widely used in actual life, as: the monitoring of all kinds of ATM cash dispensers of bank to remittee, customs, the monitoring of critical point to turnover passenger.In actual face image processing process, the blocking of facial image (as glasses, mouth mask, scarf etc.) can often occur, and block for recognition of face or face synthetic be a great obstacle.Therefore, how to detect quickly and automatically face and block and rebuild the image of face occlusion area, become one of study hotspot of face image processing in recent years.

Current existing occlusion detection technology is mainly the processing based on to reconstructed error, that is: first with training sample to there being the image blocking to be reconstructed, obtain unscreened reconstructed image, then calculate error between the two, by the size of analytical error, judge the region being blocked.Occlusion detection technology based on error analysis mainly can be divided into: the sparse coding method based on dictionary, and based on the method for error metrics, the method distributing based on error, and method based on error structure.These methods processing while blocking continuously, have a common problem: when image dimension is lower than a certain critical value or shielded area during higher than a certain percentage point, Detection accuracy can be significantly but not gently declined.For this problem, the present invention proposes a kind of new face occlusion detection method based on structuring error coding.

Summary of the invention

The invention provides a kind of face occlusion detection method based on structuring error coding that Detection accuracy is high, feasibility is good, be suitable for processing the lower or larger situation of shielded area of image dimension.

For solving the problems of the technologies described above, the technical solution used in the present invention is: the method for the face occlusion detection based on structuring error coding is provided, comprises the following steps:

Step 1: facial image data to be detected and training sample data are stretched as to column vector;

Step 2: definition error supports, and initialization;

Step 3: under minimize CD error criterion, supported by error, calculate sparse coding and the reconstructed error of facial image data to be detected to the dictionary being formed by training sample data;

Step 4: support according to reconstructed error evaluated error;

Step 5: set up and describe the aspect graph that error supports, by aspect graph and reconstructed error, evaluated error supports again;

Step 6: iterative step 3-5, obtains reconstructed error sequence and error and support sequence;

Step 7: choose Optimal error and support, and according to the set of the pixel that is blocked in Optimal error support acquisition facial image to be detected.

Further, the column vector that facial image to be detected and each training sample are stretched as in described step 1 is the column vector that the image data matrix of m × n dimension is stretched as to M=m × n dimension, and m, n are respectively line number and the columns of view data.

Further, the error in described step 2 is supported for s ∈ { 1,1} ^m, wherein s _i=-1 represents not to be blocked, s _i=1 represents to be blocked; It is that error is supported and is initialized as s that initialization error supports _i=-1, i=1 ..., M.{ 1,1} ^mrepresent the set of M dimensional vector, and in this set the element of column vector from set { 1,1}.

Further, the dictionary being made up of training sample in described step 3 is by the training sample after each stretch processing, by row discharge, forms dictionary.

Further, the CD error in described step 3 is for measuring the vector of any two same dimension with between error, be defined as: CD (a _i, b _i)=1-exp (| loga _i-logb _i| σ), wherein, θ is empirical constant, and recommended value is 1; I is the subscript of vectorial a and b, a _i, b _irepresent i the element of vectorial a and b, M is the dimension of the column vector after stretching, represent real number field middle dimension is the vectorial set of M.

Further, sparse coding and the reconstructed error of the facial image to be detected in described step 3 to the dictionary being made up of training sample calculates as follows: $(x,e)=\arg \underset{x,e}{\min }{\mathrm{\Σ}}_{i=1}^M({\stackrel{\·}{s}}_i-{\stackrel{\·\·}{s}}_i)e_{i}s.t.e_i=\mathrm{CD}(y_i,{\hat{y}}_i),\hat{y}=\mathrm{Dx},x\≥0,$ Wherein, x is non-negative sparse coding, and e is reconstructed error, ${\stackrel{\·}{s}}_i=(1-s_i)/2,{\stackrel{\·\·}{s}}_i=(1+s_i)/2,$ D is the dictionary being made up of training sample; S.t. represent that (x, e) needs to meet and the constraint of x>=0, e _irepresent y _iwith cD error.

Further, the concrete steps that in described step 4, evaluated error supports are: if iteration first, t=1, carries out two class mean clusters to reconstruct error e, that is: K mean cluster, and K=2, obtains error and supports s, and initialization threshold tau ⁽¹⁾=max{e _i| s _i=-1}; Otherwise t > 1, carries out threshold value cluster to reconstruct error e, obtain error and support $s_i=\left\{\begin{array}{cc}1,& \left|e_i\right|>{\mathrm{\τ}}^{\left(t\right)}\\ -1,& \left|e_i\right|\≤{\mathrm{\τ}}^{\left(t\right)}\end{array}\right.,$ Wherein, threshold value ${\mathrm{\τ}}^{\left(t\right)}=\left\{\begin{array}{cc}{\mathrm{\τ}}^{(t-1)}+\frac{\mathrm{\κ}-1}{T-1}{\mathrm{\τ}}^{\left(1\right)},& t\≤T\\ {\mathrm{\τ}}^{(t-1)}-\frac{\mathrm{\κ}-1}{T-1}{\mathrm{\τ}}^{\left(1\right)},& t>T\end{array}\right.,$ T and κ are empirical parameter, and t is iterations.The value suggestion of T and κ is T=5, κ=0.3.

Further, the concrete steps of described step 5 are:

Step 5.1: set up the aspect graph G=(V, E, B) that describes error support s, wherein: the set V={1 on the summit (Vertex) that V is G, 2 ..., M} and each vertex v _iclass be designated as s _i; E is the set E={ (i, j) on the limit (Edge) of G | i, and j ∈ V, || c _i-c _j|| ₂=1}, wherein c _i=[c _i1, c _i2] ^t, c _j=[c _j1, c _j2] ^tit is vertex v _i, v _jcoordinate; B is the set B={ B on the border of each subgraph of G _k| k=-1,1}, wherein: B _k=(V _k, E _k), $v_i^o=\left\{j\right|(i,j)\∈E,s_i\≠s_j\},$ $v_{\mathrm{ij}}^o=\left\{(k,l)\right|(k,l)\∈E,k\∈v_i^o,l\∈v_j^o\};$ K belongs to set l belongs to set

Step 5.2: make s'=s, by aspect graph G and reconstructed error e, evaluated error supports s again: $s=\arg \underset{s}{\max }\underset{(i,j)\∈E}{\mathrm{\Σ}}{\mathrm{\λ}}_E{s}_i{s}_j+\underset{i\∈V}{\mathrm{\Σ}}[(\log \stackrel{\·}{\mathrm{\λ}}-\stackrel{\·}{\mathrm{\λ}}{e}_i){\stackrel{\·}{s}}_i+(\log \stackrel{\·\·}{\mathrm{\λ}}-\stackrel{\·\·}{\mathrm{\λ}}+\stackrel{\·\·}{\mathrm{\λ}}{e}_i){\stackrel{\·\·}{s}}_i+{\mathrm{\λ}}_V{s}_i]-\underset{i\∈V_1}{\mathrm{\Σ}}{\mathrm{\λ}}_B{s}_i,$ Wherein, λ _efor smoothing parameter, $\stackrel{\·}{\mathrm{\λ}}={\mathrm{\Σ}}_{i\∈V}{\stackrel{\·}{s}}_i^{\′}/{\mathrm{\Σ}}_{i\∈V}{\stackrel{\·}{s}}_i^{\′}{e}_i,\stackrel{\·\·}{\mathrm{\λ}}={\mathrm{\Σ}}_{i\∈V}{\stackrel{\·\·}{s}}_i^{\′}/{\mathrm{\Σ}}_{i\∈V}{\stackrel{\·\·}{s}}_i^{\′}(1-e_i),$ λ _bfor boundary parameter.λ _erecommended value be 2; λ _brecommended value be 0.5.This optimized-type meets Ising model, can be solved by GraphCuts.

Further, the reconstructed error sequence of described step 6 is E={e ^(t)| t=1,2 ..., 2T-1}, wherein e ^(t)the reconstructed error producing for the t time iteration of above-mentioned steps 3-5; It is S={s that error supports sequence ^(t)| t=1,2 ..., 2T-1}, wherein s ^(t)the error producing for the t time iteration of above-mentioned steps 3-5 supports.

Further, the concrete steps of described step 7 are:

Step 7.1: step 7.1 makes ${\stackrel{\·}{E}}^{\left(t\right)}=(1+\log ({\mathrm{\Σ}}_{i\∈V}{\stackrel{\·}{s}}_i^{\left(t\right)}{e}_i^{\left(t\right)}/{\mathrm{\Σ}}_{i\∈V}{\stackrel{\·}{s}}_i^{\left(t\right)})){\mathrm{\Σ}}_{i\∈V}{\stackrel{\·}{s}}_i^{\left(t\right)},$ Wherein V={1,2 ..., M}; ${\stackrel{\·\·}{R}}^{\left(t\right)}={\mathrm{\Σ}}_{i\∈V_1^{\left(t\right)}}{s}_i^{\left(t\right)},$ Wherein $v_i^o=\left\{j\right|(i,j)\∈E,s_i^{\left(t\right)}\≠s_j^{\left(t\right)}\}.$

Step 7.2: order $\stackrel{\·}{E}=\left\{{\stackrel{\·}{E}}^{\left(t\right)}\right|t=1,...,2T-1\},\stackrel{\·\·}{R}=\{{\stackrel{\·\·}{R}}^{\left(t\right)}|t=1,...,2T-1\},$ To all t=1 ..., 2T-1, standardization with to interval [0,1]: ${\stackrel{\·}{E}}^{\left(t\right)}=\frac{{\stackrel{\·}{E}}^{\left(t\right)}-\min \stackrel{\·}{E}}{\max \stackrel{\·}{E}-\min \stackrel{\·}{E}},{\stackrel{\·\·}{R}}^{\left(t\right)}=\frac{{\stackrel{\·\·}{R}}^{\left(t\right)}-\min \stackrel{\·\·}{R}}{\max \stackrel{\·\·}{R}-\min \stackrel{\·\·}{R}};$

Step 7.3: to error energy carry out border regularization: wherein, ${\mathrm{\λ}}_B=\frac{{\mathrm{\Σ}}_{t=1}^{2T-2}|{\stackrel{\·}{E}}^{(t+1)}-{\stackrel{\·}{E}}^{\left(t\right)}|}{{\mathrm{\Σ}}_{t=1}^{2T-2}|{\stackrel{\·\·}{R}}^{(t+1)}-{\stackrel{\·\·}{R}}^{\left(t\right)}|};$

Step 7.4: choose optimum error and support wherein

Step 7.5: by obtain the set of the pixel being blocked of all images to be detected: $O=\left\{i\right|{\hat{s}}_i=1\},i=1,...,M.$

Compared with prior art, beneficial effect is: adopt the method for the invention, when image dimension is lower than a certain critical value or shielded area during higher than a certain percentage point, Detection accuracy can be not significantly but not is gently declined, improve accuracy rate and the scope of application to face occlusion detection, there is significant practical value.

Brief description of the drawings

Fig. 1 is the schematic flow sheet of detection method of the present invention;

Fig. 2 is the schematic diagram that image array is stretched as to column vector;

Fig. 3 is from the image to be detected (a) of AR picture library and the schematic diagram of training plan image set (b);

Fig. 4 is the sparse coding schematic diagram of image to be detected about the dictionary being made up of training sample data;

Fig. 5 is reconstructed image (a) and reconstructed error (b) schematic diagram of image to be detected about the dictionary being made up of training sample data;

Fig. 6 is by two class means clustering methods, and by reconstructed error (a), evaluated error supports the schematic diagram of (b);

Fig. 7 supports (a) by error to set up the schematic diagram of aspect graph (b), and wherein, black represents not to be blocked, and grey represents to be blocked, represent the to be blocked border in region of the white in aspect graph (b);

Fig. 8 is passing threshold clustering method, is supported the schematic diagram of (b) by reconstructed error (a) evaluated error;

Fig. 9 be by threshold value clustering method estimate error support (a), further estimate error support the schematic diagram of (b);

Figure 10 is to image to be detected and training set in Fig. 3, and the schematic diagram that all previous iteration bears results wherein, (a) is the reconstructed image of image to be detected; (b) reconstructed error sequence; (c) error supports sequence; (d) threshold series; (e) quality evaluation sequence;

Figure 11 is the testing result of the present invention to all kinds of all previous iteration of blocking, wherein, (a) be that scarf blocks, (b) be that sunglasses blocks, (c) be that monkey is blocked, (d) be that apple blocks, (a)-(b) be the blocking of real scene shooting, from AR database; (c)-and (d) be artificial synthetic blocking, original image (containing the facial image blocking) is from Extended Yale B database.

Embodiment

According to drawings and embodiments invention is described further below.

As shown in Figure 1, specific implementation step of the present invention can be expressed as follows:

1, by facial image data to be detected be stretched as column vector wherein M=m × n.

2, will come from K people's (wherein N _kbe the sample number of k people's facial image) individual not containing the facial image data of blocking (i=1 ..., N) and as training sample.Make M=m × n, by C _ibe stretched as one dimension column data vector (i=1 ..., N), form dictionary

3, initialization error supports s ∈ { 1,1} ^m(s _i=-1 represents y _ibe not blocked, s _i=1 represents y _ibe blocked): $s_i^{\left(0\right)}=-1(i=1,...,M).$

4, initialization iterations: t=0.

5, make iterations: t=t+1.

6,, under the criterion of minimize CD error, calculate the facial image y to be detected sparse coding x about dictionary D ^(t)with reconstructed error e ^(t):

$(x^{\left(t\right)},e^{\left(t\right)})=\arg \underset{x,e}{\min }{\mathrm{\Σ}}_{i=1}^M({\stackrel{\·}{s}}_i^{(t-1)}-{\stackrel{\·\·}{s}}_i^{(t-1)})e_{i}s.t.e_i=\mathrm{CD}(y_i,{\hat{y}}_i),\hat{y}=\mathrm{Dx},x\≥0---\left(1\right)$ Wherein, ${\stackrel{\·}{s}}_i^{(t-1)}=(1-s_i^{(t-1)})/2,{\stackrel{\·\·}{s}}_i^{(t-1)}=(1+s_i^{(t-1)})/2$ And $\mathrm{CD}(y_i,{\hat{y}}_i)=1-\exp (-|\log y_i-\log {\hat{y}}_i|/\mathrm{\σ}),$ Wherein, $\mathrm{\σ}=\frac{\mathrm{\θ}}{M}{\mathrm{\Σ}}_{i=1}^M{(\log y_i-\log {\hat{y}}_i)}^2,$ θ is empirical constant (determined by user, recommended value is 1).

7 if iteration first, and t=1, to reconstruct error e ⁽¹⁾carry out two class mean clusters, support s with evaluated error ⁽¹⁾, and initialization threshold tau ⁽¹⁾:

${\mathrm{\τ}}^{\left(1\right)}=\max \left\{e_i^{\left(1\right)}\right|s_i^{\left(1\right)}=-1\},---\left(2\right)$

Skip to following step 10; Otherwise t > 1, skips to following step 8.

8, estimate threshold tau ^(t):

${\mathrm{\τ}}^{\left(t\right)}=\left\{\begin{array}{cc}{\mathrm{\τ}}^{(t-1)}+\frac{\mathrm{\κ}-1}{T-1}{\mathrm{\τ}}^{\left(1\right)},& t\≤T\\ {\mathrm{\τ}}^{(t-1)}-\frac{\mathrm{\κ}-1}{T-1}{\mathrm{\τ}}^{\left(1\right)},& t>T\end{array}\right.;---\left(3\right)$

Here, T and κ are empirical parameter, and recommended value is T=5, κ=0.3.

9, by reconstructed error e ^(t)and threshold tau ^(t)evaluated error supports s ^(t):

$s_i^{\left(t\right)}=\left\{\begin{array}{cc}1& \left|e_i^{\left(t\right)}\right|>{\mathrm{\τ}}^{\left(t\right)}\\ -1& \left|e_i^{\left(t\right)}\right|\≤{\mathrm{\τ}}^{\left(t\right)}\end{array}\right..---\left(4\right)$

10, set up and describe error support s ^(t)aspect graph G ^b=(V, E, B), wherein: V is G ^bthe set V={1 on summit (Vertex), 2 ..., M} and each vertex v _iclass be designated as e is G ^bthe set E={ (i, j) on limit (Edge) | i, j ∈ V, || c _i-c _j|| ₂=1}, wherein c _i=[c _i1, c _i2] ^tit is vertex v _icoordinate; B is G ^bthe set B={ B on border of each subgraph _k| k=-1,1}, wherein: B _k=(V _k, E _k), $v_i^o=\left\{j\right|(i,j)\∈E,s_i^{\left(t\right)}\≠s_j^{\left(t\right)}\},$ $v_{\mathrm{ij}}^o=\left\{(k,l)\right|(k,l)\∈E,k\∈v_i^o,l\∈v_j^o\}.$

11, make s'=s ^(t), by aspect graph G ^bwith reconstructed error e ^(t), evaluated error supports s again ^(t): $s^{\left(t\right)}=\arg \underset{s}{\max }\underset{(i,j)\∈E}{\mathrm{\Σ}}{\mathrm{\λ}}_E{s}_i{s}_j+\underset{i\∈V}{\mathrm{\Σ}}[(\log \stackrel{\·}{\mathrm{\λ}}-\stackrel{\·}{\mathrm{\λ}}{e}_i^{\left(t\right)}){\stackrel{\·}{s}}_i+(\log \stackrel{\·\·}{\mathrm{\λ}}-\stackrel{\·\·}{\mathrm{\λ}}+\stackrel{\·\·}{\mathrm{\λ}}{e}_i^{\left(t\right)}){\stackrel{\·\·}{s}}_i+{\mathrm{\λ}}_V{s}_i]-\underset{i\∈V_1}{\mathrm{\Σ}}{\mathrm{\λ}}_B{s}_i,---\left(5\right)$ Wherein: λ _efor smoothing parameter (given by user, recommended value is 2), λ _bfor boundary parameter (determined by user, recommended value is 0.5).This optimized-type meets Ising model, can be solved by GraphCuts.

12, calculate the error energy that does not block part

${\stackrel{\·}{E}}^{\left(t\right)}=(1+\log ({\mathrm{\Σ}}_{i\∈V}{\stackrel{\·}{s}}_i^{\left(t\right)}{e}_i^{\left(t\right)}/{\mathrm{\Σ}}_{i\∈V}{\stackrel{\·}{s}}_i^{\left(t\right)})){\mathrm{\Σ}}_{i\∈V}{\stackrel{\·}{s}}_i^{\left(t\right)}---\left(6\right)$

With the regular coefficient that blocks border partly

${\stackrel{\·\·}{R}}^{\left(t\right)}={\mathrm{\Σ}}_{i\∈V_1}{s}_i^{\left(t\right)}.---\left(7\right)$

13, iteration above-mentioned steps 4～12, until maximum iteration time 2T-1.

14, order $\stackrel{\·}{E}=\left\{{\stackrel{\·}{E}}^{\left(t\right)}\right|t=1,...,2T-1\},\stackrel{\·\·}{R}=\{{\stackrel{\·\·}{R}}^{\left(t\right)}|t=1,...,2T-1\},$ To all t=1 ..., 2T-1, standardization with to interval [0,1]:

${\stackrel{\·}{E}}^{\left(t\right)}=\frac{{\stackrel{\·}{E}}^{\left(t\right)}-\min \stackrel{\·}{E}}{\max \stackrel{\·}{E}-\min \stackrel{\·}{E}},{\stackrel{\·\·}{R}}^{\left(t\right)}=\frac{{\stackrel{\·\·}{R}}^{\left(t\right)}-\min \stackrel{\·\·}{R}}{\max \stackrel{\·\·}{R}-\min \stackrel{\·\·}{R}};---\left(8\right)$

To error energy carry out border regularization:

$C^{\left(t\right)}={\stackrel{\·}{E}}^{\left(t\right)}-{\mathrm{\λ}}_B{\stackrel{\·\·}{R}}^{\left(t\right)}---\left(9\right)$

Wherein, ${\mathrm{\λ}}_B=\frac{{\mathrm{\Σ}}_{t=1}^{2T-2}|{\stackrel{\·}{E}}^{(t+1)}-{\stackrel{\·}{E}}^{\left(t\right)}|}{{\mathrm{\Σ}}_{t=1}^{2T-2}|{\stackrel{\·\·}{R}}^{(t+1)}-{\stackrel{\·\·}{R}}^{\left(t\right)}|}.$

15, choosing optimum error supports wherein by obtain the set of all pixels that are blocked:

For better explanation, provide concrete example below:

1, from AR storehouse, get and in the 2nd people, have sample that scarf blocks as image y to be detected, dimension is 112 × 92, it is stretched as to 10304 × 1 column vector (10304=112 × 92) according to the mode shown in Fig. 2.32 that from AR storehouse, get front 4 people do not block sample (everyone 8 samples) as training sample set, as shown in Figure 3, the dimension of each sample is 112 × 92, by each sample according to shown in Fig. 2 mode be stretched as 10304 × 1 column vector, and the training sample set after stretching is organized as to dictionary D, now the dimension of dictionary D is 10304 × 32.

2, initialization error supports s ⁽⁰⁾(dimension is 10304 × 1), determine parameter: λ _e=2, λ _b=0.5, T=5, κ=0.3.

3, make t=1, image y to be detected, initialization error are supported to s ⁽⁰⁾with training sample set D, substitution formula (1), calculate y about the sparse coding x of D ⁽¹⁾with reconstructed error e ⁽¹⁾, as shown in Figure 4, Figure 5.To reconstruct error e ^(t)carry out the mean cluster of two classes, evaluated error supports s ⁽¹⁾=K (e ⁽¹⁾), as shown in Figure 6, black represents (not being blocked), grey represents (being blocked), by formula (2) initialization threshold tau ⁽¹⁾=0.4886.

4, set up and characterize error support s ⁽¹⁾aspect graph G ^b=(V, E, B), as shown in Figure 7, black represents (not being blocked), grey represents (being blocked), white represents to be blocked the boundary B in region ₁.Make s'=s ⁽¹⁾, s ⁽¹⁾, e ⁽¹⁾, G ^bsubstitution formula (5), reappraises s ⁽¹⁾.

5,, by formula (6) and (7), calculate the error energy that does not block part with the regular coefficient that blocks border partly

6, make t=2 that image y to be detected, initialization error are supported to s ⁽¹⁾with sample training collection D, substitution formula (1), calculate y about the non-negative sparse coding x of D ⁽²⁾with reconstructed error e ⁽²⁾, estimate threshold tau by formula (3) ⁽²⁾=0.4031.

7, reconstructed error e ⁽²⁾and threshold tau ⁽²⁾substitution formula (4), evaluated error supports s ⁽²⁾, as shown in Figure 8, passing threshold clustering method, is supported by reconstructed error evaluated error.

8, set up and describe error support s ⁽²⁾aspect graph G ^b=(V, E, B), makes s'=s ⁽²⁾, $\stackrel{\·}{\mathrm{\λ}}={\mathrm{\Σ}}_{i\∈V}{\stackrel{\·}{s}}_i^{\′}/{\mathrm{\Σ}}_{i\∈V}{\stackrel{\·}{s}}_i^{\′}{e}_i^{\left(t\right)}=4.2,\stackrel{\·\·}{\mathrm{\λ}}={\mathrm{\Σ}}_{i\∈V}{\stackrel{\·\·}{s}}_i^{\′}/{\mathrm{\Σ}}_{i\∈V}{\stackrel{\·\·}{s}}_i^{\′}(1-e_i^{\left(t\right)})=21.9,$ S ⁽²⁾, e ⁽²⁾, G ^bsubstitution formula (5), reappraises s ⁽²⁾.Fig. 9 be by threshold value clustering method estimate error support, further estimate error support, wherein black is not blocked, grey represents to be blocked.

9,, by formula (6) and (7), calculate the error energy that does not block part with the regularity of blocking border partly

10, after 2T-1=9 iteration, obtain with sequence, as Table I:

The all previous iteration of Table I calculate with

11, right by formula (8) with standardize, as Table II:

After Table II standardization with

12, by formula (9) to error energy carry out border regularization, as shown in Table III.

Table III by formula (9) to error energy carry out border regularization (λ _b=1.2726)

13, choosing optimum error supports (seeing last row of Figure 10 (b)), wherein $\hat{t}=\arg {\min }_t{C}^{\left(t\right)}=9.$

14, by obtain the set of all pixels that are blocked: as the grey part in the picture of last row of Figure 10 (b).

Figure 10 is to image to be detected and training set in Fig. 3, and the schematic diagram that all previous iteration bears results wherein, (a) is the reconstructed image of image to be detected; (b) reconstructed error sequence; (c) error supports sequence, and black is not blocked, and grey represents to be blocked; (d) threshold series; (e) quality evaluation sequence, assessed value 0.3101 minimum (being risen by square frame frame), its corresponding testing result, in (c), the error of last row supports, also best, supports s with the error of other each iterative estimate that is: ^(t)(t=1,2 ..., 8) compare, estimation be optimum, so by the pixel being blocked obtaining is also close to truth, and effect is better.

Figure 11 is the testing result of the present invention to all kinds of all previous iteration of blocking, wherein, (a) be that scarf blocks, (b) be that sunglasses blocks, (c) be that monkey is blocked, (d) be that apple blocks, (a)-(b) be the blocking of real scene shooting, from AR database; (c)-and (d) be artificial synthetic blocking, original image (containing the facial image blocking) is from Extended Yale B database; S supports sequence corresponding to error, and black represents not to be blocked, and grey represents to be blocked, and white represents the edge of occlusion area; τ is corresponding to threshold series; C is corresponding to quality evaluation sequence, minimum value with square frame frame get up, it is the result detecting that its corresponding error supports.The optimal detection result that can see each example occurs at different iterationses, need to choose optimum result by quality evaluation.

Claims (1)

1. the face occlusion detection method based on structuring error coding, is characterized in that, comprises the following steps:

Step 1: facial image data to be detected and training sample data are stretched as to column vector;

Step 2: definition error supports, and initialization;

Step 3: under minimize CD error criterion, supported by error, calculate sparse coding and the reconstructed error of facial image data to be detected to the dictionary being formed by training sample data;

Step 4: support according to reconstructed error evaluated error;

Step 5: set up and describe the aspect graph that error supports, by aspect graph and reconstructed error, evaluated error supports again;

Step 6: iterative step 3-5, obtains reconstructed error sequence and error and support sequence;

Step 7: choose Optimal error and support, and according to the set of the pixel that is blocked in Optimal error support acquisition facial image to be detected;

Column vector that facial image to be detected and each training sample are stretched as in described step 1 is the column vector that the image data matrix of m ' n dimension is stretched as to M=m ' n dimension;

Error in described step 2 is supported for s ∈ { 1,1} ^m, wherein s _i=-1 represents not to be blocked, s _i=1 represents to be blocked; It is error to be supported to s be initialized as s that initialization error supports _i=-1, i=1 ..., M; I is the subscript of s, s _irepresent i the element of s;

The dictionary being made up of training sample in described step 3 is by the training sample after each stretch processing, by row discharge, forms dictionary;

CD error in described step 3 is for measuring the vector of any two same dimension with between error, be defined as: CD (a _i, b _i)=1-exp (| loga _i-logb _i|/σ), wherein, q is empirical constant; I is the subscript of vectorial a and b, a _i, b _irepresent i the element of a and b, M is the dimension of the column vector after stretching, represent real number field middle dimension is the vectorial set of M;

Sparse coding and the reconstructed error of facial image to be detected in described step 3 to the dictionary being made up of training sample calculates as follows: $(x,e)=\arg \underset{x,e}{\min }{\mathrm{\Σ}}_{i=1}^M({\stackrel{\·}{s}}_i-{\stackrel{\·\·}{s}}_i)e_{i}s.t.e_i=\mathrm{CD}(y_i,{\hat{y}}_i),\hat{y}=\mathrm{Dx},x\≥0,$ Wherein, x is non-negative sparse coding, and e is reconstructed error, d is the dictionary being made up of training sample, and y is facial image to be detected; S.t. represent that (x, e) needs to meet with the constraint of x30, e _irepresent y _iwith cD error;

The concrete steps that in described step 4, evaluated error supports are: if iteration first, t=1, carries out two class mean clusters to reconstruct error e, obtain error and support s, and initialization threshold value t ⁽¹⁾=max{e _i| s _i=-1}; Otherwise t>1, carries out threshold value cluster to reconstruct error e, obtain error and support $s_i=\left\{\begin{array}{c}1,\left|e_i\right|>{\mathrm{\τ}}^{\left(t\right)}\\ -1,\left|e_i\right|\≤{\mathrm{\τ}}^{\left(t\right)}\end{array}\right.,$ Wherein, threshold value ${\mathrm{\τ}}^{\left(t\right)}=\left\{\begin{array}{c}{\mathrm{\τ}}^{(t-1)}+\frac{\mathrm{\κ}-1}{T-1}{\mathrm{\τ}}^{\left(1\right)},t\≤T\\ {\mathrm{\τ}}^{(t-1)}-\frac{\mathrm{\κ}-1}{T-1}{\mathrm{\τ}}^{\left(1\right)},t>T\end{array}\right.,$ T and k are empirical parameter, and t is iterations;

Described step 5 concrete steps are:

Step 5.1: set up the aspect graph G=(V, E, B) that describes error support s, wherein: the set V={1 on the summit that V is G, 2 ..., M} and each vertex v _iclass be designated as s _i; E is the set E={ (i, j) on the limit of G | i, and j ∈ V, || c _i-c _j|| ₂=1}, wherein c _i=[c _i1, c _i2] ^t, c _j=[c _j1, c _j2] ^tit is vertex v _i, v _jcoordinate; B is the set B={ B on the border of each subgraph of G _k| k=-1,1}, wherein: B _k=(V _k, E _k), $v_{\mathrm{ij}}^o=\left\{(k,l)\right|(k,l)\∈E,k\∈v_i^0,l\∈v_j^0\};$ K belongs to set l belongs to set

Step 5.2: make s'=s, by aspect graph G and reconstructed error e, evaluated error supports s again: $s=\arg \underset{s}{\max }\underset{(i,j)\∈E}{\mathrm{\Σ}}{\mathrm{\λ}}_E{s}_i{s}_j+\underset{i\∈V}{\mathrm{\Σ}}[(\log \stackrel{\·}{\mathrm{\λ}}-\stackrel{\·}{\mathrm{\λ}}{e}_i){\stackrel{\·}{s}}_i+(\log \stackrel{\·\·}{\mathrm{\λ}}-\stackrel{\·\·}{\mathrm{\λ}}+\stackrel{\·\·}{\mathrm{\λ}}{e}_i){\stackrel{\·\·}{s}}_i+{\mathrm{\λ}}_V{s}_i]-\underset{i\∈V_1}{\mathrm{\Σ}}{\mathrm{\λ}}_B{s}_i,$ Wherein, λ _efor smoothing parameter, $\stackrel{\·}{\mathrm{\λ}}={\mathrm{\Σ}}_{i\∈V}{\stackrel{\·}{s}}_i^{\′}/{\mathrm{\Σ}}_{i\∈V}{\stackrel{\·}{s}}_i^{\′}{e}_i,\stackrel{\·\·}{\mathrm{\λ}}={\mathrm{\Σ}}_{i\∈V}{\stackrel{\·\·}{s}}_i^{\′}/{\mathrm{\Σ}}_{i\∈V}{\stackrel{\·\·}{s}}_i^{\′}(1-e_i),$ λ _bfor boundary parameter;

The reconstructed error sequence of described step 6 is E={e ^(t)| t=1,2 ..., 2T-1}, wherein e ^(t)the reconstructed error producing for the t time iteration of step 3-5; It is S={s that error supports sequence ^(t)| t=1,2 ..., 2T-1}, wherein s ^(t)the error producing for the t time iteration of step 3-5 supports;

The concrete steps of described step 7 are:

Step 7.1: order ${\stackrel{\·}{E}}^{\left(t\right)}=(1+\log ({\mathrm{\Σ}}_{i\∈V}{\stackrel{\·}{s}}_i^{\left(t\right)}{e}_i^{\left(t\right)}/{\mathrm{\Σ}}_{i\∈V}{\stackrel{\·}{s}}_i^{\left(t\right)})){\mathrm{\Σ}}_{i\∈V}{\stackrel{\·}{s}}_i^{\left(t\right)},$ Wherein V={1,2 ..., M}; wherein

Step 7.2: order $\stackrel{\·}{E}=\left\{{\stackrel{\·}{E}}^{\left(t\right)}\right|t=1,...,2T-1\}$ , to all t=1 ..., 2T-1, standardization with to interval [0,1]: ${\stackrel{\·}{E}}^{\left(t\right)}=\frac{{\stackrel{\·}{E}}^{\left(t\right)}-\min \stackrel{\·}{E}}{\max \stackrel{\·}{E}-\min \stackrel{\·}{E}},{\stackrel{\·\·}{R}}^{\left(t\right)}=\frac{{\stackrel{\·\·}{R}}^{\left(t\right)}-\min \stackrel{\·\·}{R}}{\max \stackrel{\·\·}{R}-\min \stackrel{\·\·}{R}};$

Step 7.3: to error energy carry out border regularization: $C^{\left(t\right)}={\stackrel{\·}{E}}^{\left(t\right)}-{\mathrm{\λ}}_B{\stackrel{\·\·}{R}}^{\left(t\right)},$ Wherein, ${\mathrm{\λ}}_B=\frac{{\mathrm{\Σ}}_{t=1}^{2T-2}|{\stackrel{\·}{E}}^{(t+1)}-{\stackrel{\·}{E}}^{\left(t\right)}|}{{\mathrm{\Σ}}_{t=1}^{2T-2}|{\stackrel{\·\·}{R}}^{(t+1)}-{\stackrel{\·\·}{R}}^{\left(t\right)}|};$

Step 7.4: choose optimum error and support wherein by obtain the set of the pixel being blocked of all images to be detected: i=1 ..., M.