CN107066943B - A face detection method and device - Google Patents

Abstract

The invention discloses a face detection method and device. The method of the present invention is: 1) detect the candidate face from the image to be processed, and extract the candidate features of the candidate face; The feature space is projected and transformed to obtain the corresponding traditional or approximate embedded features; wherein, the approximate external feature space is a dictionary composed of representative features selected from the reference face feature dictionary and non-face feature dictionary; 3) for all The embedding feature is verified to determine whether the candidate face corresponding to the embedding feature is a human face. The face detection device of the present invention includes a candidate module, an embedding module and a verification module. The present invention can obtain higher precision human face detection performance, and also has good human face detection ability in the case of occlusion.

Description

A face detection method and device

technical field

The invention belongs to the field of computer vision and deep learning, and in particular relates to a face detection method and device under occlusion conditions.

Background technique

Face detection technology can be applied to camera autofocus, human-computer interaction, photo management, urban security monitoring, intelligent driving and many other fields. At present, in the practical application of face detection in open environment conditions, due to the ubiquity of occlusion (such as dense crowds), the performance of face detection is seriously challenged, so the performance of face detection under occlusion conditions needs to be solved. In addition, it is of great practical significance to study face detection under masked conditions, for example: in video surveillance, it is used to find suspicious persons to provide warnings, and to predict weather conditions by detecting the distribution of masked faces. Traditional face detection methods suffer from severe performance degradation under occlusion, because the face clues of the occluded part are invalid during the detection process, which inevitably introduces noise in the feature extraction process. In conclusion, incomplete and inaccurate features make masked face detection an extremely challenging problem.

In recent years, some methods have also been studied in this field. The existing technology is to detect face candidates first, and then classify and confirm the face candidates. One of the methods detects face candidates by training multiple neural networks to obtain the responses of multiple parts of the face, and then trains a new neural network to classify and confirm face candidates (see: S.Yang, P.Luo, C. C. Loy, and X. Tang. From facial parts responses to face detection: A deep learning approach. In: IEEE ICCV, 2015). Another method is to confirm the face candidate by selecting some feature comparisons to calculate the loss (see: M.Opitz, G.Waltner, G.Poier, H.Possegger, and H.Bischo. Grid loss: Detecting occluded faces.In ECCV , 2016), this method can better deal with the face detection problem of partial occlusion. The above method alleviates the problem of face detection in the case of severe occlusion (such as mask occlusion) to a certain extent, but it still cannot be completely solved. When the face parts are occluded, the face detection method that detects face candidates through the response of multiple parts will introduce noise or errors in the parts in the occluded area, which may lead to errors in face classification confirmation; when the occlusion is serious, The face detection method that confirms the face candidate by selecting some features and comparing the calculation loss to confirm the face candidate, the calculated loss has a large error, which leads to the failure of face detection.

Contents of the invention

In order to overcome the deficiencies in the prior art, the present invention provides a face detection method and device, which detects candidate faces and extracts high-dimensional depth features (i.e. candidate features) through convolutional neural networks, and then performs local linear embedding. Feature projection is used to eliminate the incomplete and inaccurate features caused by mask occlusion, and then the multi-task convolutional neural network (ie CNN-V) is used to verify the candidate face, so as to obtain more accurate face detection performance. At the same time, the present invention also proposes a method for constructing an approximate external feature space, by searching for the most similar reference face and the most different reference non-face from an external database, and constructing an approximate external feature space, using the approximate external feature Embedding transforms the candidate features in space to modify the candidate features. The present invention is realized through the following technical solutions.

A kind of human face detection method of the present invention, its step comprises:

1) Perform candidate face detection on the image to be detected to obtain a candidate face image;

2) Carry out candidate feature extraction to described candidate face image, obtain candidate feature;

3) performing embedding transformation on the candidate features to obtain traditional embedding features or approximate embedding features, the embedding features can restore face clues and remove noise caused by occlusion;

4) The traditional embedded features or approximate embedded features are verified through classification and regression algorithms to obtain detection results.

Further, the candidate features are embedded and transformed through a pre-built external feature space to obtain traditional embedded features or approximate embedded features; the external feature space is traditional external feature space or approximate external feature space.

Further, the embedding transformation is realized by the traditional local linear embedding method or the fast approximate local linear embedding method; the traditional local linear embedding method uses the traditional external feature space to perform embedding transformation on the candidate features with noise to obtain the traditional embedded features; the fast approximate local linear embedding method Linear embedding uses the approximate external feature space to perform embedding transformation on the candidate features with noise to obtain approximate embedded features.

Further, the method for constructing an approximate external feature space in the fast approximate local linear embedding method includes the following steps:

a) Perform candidate face detection and candidate feature extraction on the marked reference face dataset, judge whether the candidate features belong to human face features or non-human face features, and store these candidate features in the reference face feature dictionary and reference non-human feature dictionary respectively face feature dictionary;

b) Perform candidate face detection and candidate feature extraction on the marked masked face dataset, determine whether the candidate features belong to masked face features or masked non-face features, and store these candidate features into masked faces Feature dictionary and masked non-face feature dictionary;

c) select a representative reference face feature dictionary that can represent the above-mentioned masked face feature dictionary from the above-mentioned reference face feature dictionary;

D) select a representative reference non-face feature dictionary that can represent the above-mentioned masked non-face feature dictionary from the above-mentioned reference non-face feature dictionary;

e) Merge the above-mentioned representative reference face feature dictionary and representative reference non-face feature dictionary to obtain an approximate external feature space.

Further, in step a), it is determined by calculating the degree of overlap between the candidate face position corresponding to the candidate feature and the marked face position, and the degree of overlap is measured by the intersection ratio, wherein the intersection ratio is greater than If it is 0.7, it is judged that the candidate feature is the feature of the reference face, and if the intersection ratio is less than 0.3, it is judged that the candidate feature is the feature of the reference non-face.

Further, in step b), it is determined by calculating the degree of overlap between the candidate face position corresponding to the candidate feature and the marked face position, and the degree of overlap is measured by the intersection ratio, wherein the intersection ratio is greater than If it is 0.6, it is judged that the candidate feature is a masked face feature, and if the intersection ratio is less than 0.4, it is judged that the candidate feature is a masked non-face feature.

Further, in step c), adopt greedy algorithm to select representative reference face feature dictionary from reference face feature dictionary; Described greedy algorithm refers to calculating the loss of each reference face feature in the reference face feature dictionary, Obtain the list of reference human face features arranged in ascending order by loss, and get the reference human face feature at the top of the list to represent the masked human face feature; The distance between the nearest neighbor feature of the face feature dictionary and the distance between each reference face feature and the nearest neighbor feature of the masked non-face feature dictionary.

Further, in step d), adopt greedy algorithm to select representative reference non-face feature dictionary from reference non-face feature dictionary; The loss of features, get the list of reference non-face features arranged in ascending order according to the loss, take the reference non-face features at the top of the list to represent masked non-face features; wherein the loss refers to each reference The difference between the distance of the non-face feature to the nearest neighbor feature of the masked non-face feature dictionary and the distance of each reference non-face feature to the nearest neighbor feature of the masked face feature dictionary.

The invention also relates to a face detection device, which includes a candidate module, an embedding module and a verification module. The candidate module is used to detect candidate faces of the image to be detected and extract candidate features; the embedding module is used to embed and transform the candidate features to obtain traditional embedded features or approximate embedded features. The embedded features can restore face clues and remove occlusions. The noise; the verification module is used to verify the traditional embedded features or approximate embedded features through classification and regression algorithms to obtain the final detection results. The candidate module obtains multiple candidate features, and then performs embedding transformation through a pre-built external feature space in the embedding module to obtain traditional embedded features or approximate embedded features; the external feature space is traditional external feature space or approximate external feature space; The embedding transformation is realized by traditional local linear embedding method or fast approximate local linear embedding method.

The beneficial effects of the present invention are:

Aiming at the face detection problem under occlusion conditions, especially the face detection problem under severe mask occlusion conditions, the detection method and device of the present invention have relatively good performance; The face detection method and device also have good processing capabilities.

Description of drawings

Fig. 1 is the flowchart of a kind of face detection method of the present invention;

Fig. 2 is a schematic flow diagram of a candidate module of the device of the present invention;

Fig. 3 is a schematic flow chart of the embedding module of the device of the present invention;

Fig. 4 is a schematic flow chart of the verification module of the device of the present invention;

Fig. 5 is a schematic diagram of the construction process of the approximate external feature space of the present invention.

Detailed ways

In order to make the above solutions and beneficial effects of the present invention more comprehensible, the following will be described in detail through the examples and accompanying drawings.

The present invention provides a kind of human face detection method and device, and this device comprises candidate module, embedding module and verification module; The flowchart of this method is as shown in Figure 1, and its steps include:

1) Receive an image. The image can be a face image under occlusion conditions or a face image under severe mask occlusion conditions, or a face image without occlusion, or an image without a face.

2) Detect the candidate face through the candidate module and extract the high-dimensional deep feature of the candidate face, that is, the candidate feature.

In the candidate module, the candidate face detection is performed first, and then it is judged whether the candidate face is detected, and if the candidate face is not detected, it ends; if the candidate face is detected, the candidate feature extraction is performed to obtain the candidate feature.

Please refer to Figure 2, the candidate module mainly includes two convolutional neural networks: one is a small convolutional neural network (called candidate convolutional neural network, referred to as CNN-P), which is used to implement candidate face detection ; Another large convolutional neural network (called feature convolutional neural network, referred to as CNN-F) is used to implement candidate feature extraction. First, the received image passes through the candidate convolutional neural network for candidate face detection, and then judges whether the candidate face is detected. If the candidate face is not detected, it ends; if the candidate face is detected, the candidate face is detected first. The face is normalized, and then the candidate features are extracted through the feature convolutional neural network to obtain the candidate features.

3) Embedding candidate features through the embedding module to obtain features after embedding transformation, that is, traditional embedding features or approximate embedding features (collectively referred to as embedding features).

Mask occlusion will cause missing face clues and feature noise, resulting in incomplete and inaccurate features. To solve this problem, the embedding module in the technical solution of the present invention realizes recovering face clues and removing noise from candidate features. The advantage of embedding module processing is that the obtained embedding features can well represent the masked face, which can improve the detection accuracy.

Please refer to Figure 3. In the embedding module, the candidate features pass through a pre-built external feature space, and after embedding transformation, traditional embedded features or approximate embedded features are obtained. The embedding transformation is mainly implemented by the LLE (LocalLinear Embedding) method. LLE is a dimensionality reduction method for nonlinear data. The processed low-dimensional data can maintain the original topological relationship. It has been widely used in the classification and clustering of image data, the visualization of multidimensional data, and bioinformatics. The invention uses the traditional LLE method and the fast approximate LLE method to realize the embedded transformation.

4) Through the verification module, the traditional embedded feature or approximate embedded feature is verified, and it is judged whether the candidate face corresponding to each traditional embedded feature or approximate embedded feature belongs to a real face, if the candidate face corresponding to the traditional embedded feature or approximate embedded feature If the face belongs to the real face, record the face information; if the candidate face corresponding to the traditional embedded feature or the approximate embedded feature does not belong to the real face, then end.

Please refer to Figure 4. The verification module consists of a four-layer fully connected convolutional neural network (called a verification convolutional neural network, or CNN-V for short), which is used for feature verification, that is, to distinguish the traditional embedded features or approximate embedded features. Whether the candidate face corresponding to the feature belongs to the real face and correct the position and scale of the corresponding candidate face. If it does not belong to a real face, ignore the candidate face corresponding to the traditional embedded feature or the approximate embedded feature; if it belongs to the real face, then the corrected candidate face position and scale corresponding to the traditional embedded feature or the approximate embedded feature added to the test results.

Through the verification module, the traditional embedded features or approximate embedded features are classified and regressed, so as to determine whether the candidate is a real face or a non-face, and correct the position and scale of the face, so as to obtain a face detection performance with higher accuracy .

Therefore, a face detection method and device proposed by the present invention combine the candidate convolutional neural network CNN-P of the candidate module, the feature convolutional neural network CNN-F of the candidate module, the verification convolutional neural network of the embedding module and the verification module. Network CNN-V, to achieve the purpose of the present invention.

The method adopted for the embedding transformation of the embedding module is described in detail below.

1. The traditional LLE method.

Please refer to Figure 3. Through the traditional LLE method, the masked candidate features x _i are projected and transformed in the pre-constructed traditional external feature space to obtain embedded features v _{i ,} which can effectively eliminate the The incomplete and inaccurate features caused by surface occlusion have good anti-occlusion capabilities. Among them, the subscript _i of xi is used to mark different candidate features; the subscript i of v _i is used to mark different embedded features. Embedded features v _i are called traditional embedded features.

The traditional external feature space is composed of reference face features and reference non-face features, which are expressed in the form of a feature dictionary, that is, D=[D _⁺ , ^D- ], where D ⁺ is a reference face feature dictionary, and ^D- It is a reference to the non-face feature dictionary. Generally speaking, the D ⁺ and ^D- scales have millions.

The reference face features and reference non-face features are realized by constructing a reference candidate feature set. Specifically, for the labeled large-scale unoccluded reference face dataset S _n , the candidate module is used to detect candidate faces and extract candidate features. Determine whether the candidate features belong to face features or non-face features, divide these candidate features into reference face features and reference non-face features, and store them in the reference face feature dictionary D ⁺ and the reference non-face feature dictionary D ⁻ respectively. Among them, judging whether the candidate feature is a face feature or a non-face feature is determined by calculating the degree of overlap between the candidate face position corresponding to the candidate feature and the marked face position, and the overlap degree is determined by the intersection ratio (Intersection -over-Union, IoU) to measure. Usually, in the traditional method, if the intersection ratio is greater than 0.5, it is judged as a human face, and if it is less than 0.5, it is judged as a non-human face. Compared with the traditional method, in the present invention, the intersection and union ratio greater than 0.7 is judged as a reference face, and the intersection and union ratio is less than 0.3 is judged as a reference non-human face, so that the obtained reference face and the reference non-human face have a better distinction It can ensure that the reference candidate features have better recognition ability.

For each noisy candidate feature x _i , select the feature set closest to x _i from D ⁺ and D ^- to form a feature subdictionary D _i (the subscript i of D _i is used to mark the corresponding feature sub-dictionary), and then use the LLE algorithm for projection transformation to obtain a new feature expression, that is, the traditional embedded feature v _i , the solution formula of this process is as follows:

satisfy v _i ≥ 0(1)

2. Fast approximate LLE method.

The present invention proposes a fast approximate LLE method. For each noisy candidate feature x _i , use the fast approximate LLE method to perform projection transformation to obtain an approximate embedded feature The solution formula of this method is as follows:

Satisfy

In the above formula (2), is an approximate external feature space, which is a dictionary composed of representative features selected from the reference face feature dictionary D ⁺ and non-face feature dictionary D ⁻ . For each candidate feature _xi, it is no longer necessary to construct its corresponding feature sub-dictionary D _i , and each candidate feature _xi uses a fixed approximate external feature space Perform projection transformation to obtain approximate embedded features

The construction of the approximate external feature space in the fast approximate LLE method is described in detail below.

The construction method of the approximate external feature space is to construct the approximate external feature space by finding the most similar reference human face or the most different reference non-human face from an external database.

Please refer to Figure 5, which is an approximate external feature space Constructed flow chart, is to select the most representative feature composition from D ⁺ and ^D- , which includes a representative reference face feature dictionary and a representative reference non-face feature dictionary Expressed as The approximate external feature space proposed by the present invention The construction method is divided into the following steps:

1) Construct reference face and reference non-face feature dictionaries: it is the same as in the above-mentioned traditional LLE method. For the marked large-scale unoccluded reference face dataset S _n , use the candidate module to detect candidate faces and candidate faces. feature extraction. According to whether the candidate features belong to face features or non-face features, these candidate features are respectively stored in the reference face feature dictionary D ⁺ and the reference non-face feature dictionary D ⁻ . Judging whether a candidate feature belongs to a face feature or a non-face feature is determined by calculating the degree of overlap between the candidate face position corresponding to the candidate feature and the marked face position, and the degree of overlap is measured by the intersection-over-union ratio (IoU) . Usually, in the traditional method, if the intersection ratio is greater than 0.5, it is judged as a human face, and if it is less than 0.5, it is judged as a non-human face. Compared with the traditional method, in the present invention, the intersection and union ratio greater than 0.7 is judged as a reference face, and the intersection and union ratio is less than 0.3 is judged as a reference non-human face, so that the obtained reference face and the reference non-human face have a better distinction It can ensure that the reference candidate features have better recognition ability.

2) Construct masked face and masked non-face feature dictionaries: similar to the above step 1), for the marked large-scale masked face dataset S _m , use the candidate module to detect candidate faces and extract candidate features. According to whether the candidate features belong to masked face features or masked non-face features, these candidate features are divided into masked face feature dictionary and a dictionary of masked non-face features Since the positioning accuracy of masked face detection is usually lower than that of unoccluded face detection, in the present invention, an intersection ratio greater than 0.6 is judged as a masked face, and an intersection ratio less than 0.4 is judged as a masked non-human face. To select masked face candidate features with better quality.

3) Select a representative reference face feature dictionary Select from the reference face feature dictionary D ⁺ , which is a subset of D ⁺ namely The representativeness of σ indicates that it has good representational power when representing masked faces and discriminative power when representing masked non-human faces. thereby, In a sparsely represented masked face feature dictionary should have minimal error while sparsely representing masked non-face feature dictionaries There should be maximum error. therefore, By solving the following formula (3), we get:

Satisfy

The above formula (3) belongs to sparse coding processing, and α ₁ and α ₂ in the formula are respectively Represents the sparse coefficient vector required for a certain masked face feature x1 and _a certain masked non-face feature _x2 . Only one element in the sparse coefficient vector is 1, and the other elements are 0. Using the constraints of sparse coefficient vectors, the sparse coding process is equivalent to starting from Find the nearest neighbor in . because Each feature in comes from the reference face feature dictionary D ⁺ . The optimization problem of formula (3) is different from the classical sparse coding method, and it is difficult to solve it with the classic optimization algorithm. Therefore, the present invention proposes a greedy method to construct effectively from the reference face feature dictionary D ⁺ In the proposed greedy method, the present invention firstly calculates each reference face feature in the reference face feature dictionary D ⁺ Loss This loss is expressed as Dictionary with masked face features The distance of the nearest neighbor feature and Dictionary with masked non-face features The distance difference of the nearest neighbor feature of , which is realized by the following formula (4):

Satisfy

In the above formula (4), ρ ₁ and ρ ₂ are two balance coefficients, which are used to balance the distance between features. In actual processing, 1 is usually used to speed up the calculation. Each reference face feature rarely used to represent Masked face features in and Masked non-face features in . Calculate the loss by Obtain a list of reference face features arranged in ascending order of loss. The reference face features at the top of the list have the strongest ability to represent masked face features, while the ability to represent masked non-face features the weakest. In this way, the top M reference face features of the list can be continuously added to a feature pool P ₊ in an iterative manner to construct the final Preferably, M is greater than or equal to 1 and less than or equal to 50. Specifically, let the initial feature pool be empty or Then at step t use To select the top M candidates, get then, The features in are used to update Then used to solve the objective function in formula (3).

4) Select a representative reference non-face feature dictionary Select from the reference non ^- face feature dictionary D-, which is a subset of ^D- The representativeness of σ indicates that it has good representational power when representing masked non-human faces and discriminative power when representing masked human faces. thereby, Sparsely represent masked non-face feature dictionaries should have minimal error while sparsely representing masked face feature dictionaries should have the greatest error. therefore, can be obtained by solving the following formula (5):

Satisfy

The above formula (5) belongs to sparse coding processing, and α ₁ and α ₂ in the formula are respectively Represents the sparse coefficient vector required for a certain masked face feature x1 and _a certain masked non-face feature _x2 . Only one element in the sparse coefficient vector is 1, and the other elements are 0. Using the constraints of sparse coefficient vectors, the sparse coding process is equivalent to starting from Find the nearest neighbor in . because Each feature in comes from the reference non-face feature dictionary D ^- , the optimization problem of formula (5) is different from the classical sparse coding method, and it is difficult to solve it with the classical optimization algorithm. Therefore, the present invention proposes a greedy method to construct effectively from the reference non ^- face feature dictionary D- In the proposed greedy method, the present invention first calculates each reference non-face feature in the reference non ^- face feature dictionary D- Loss This loss is expressed as Dictionary with masked non-face features The distance of the nearest neighbor feature and Dictionary with masked face features The distance difference between the nearest neighbor features of is realized by the following formula (6): Satisfy

In the above formula (6), ρ ₁ and ρ ₂ are two balance coefficients, which are used to balance the distance between features. In actual processing, they usually take 1 to speed up the calculation. Each reference non-face feature rarely used to represent Masked face features in and Masked non-face features in . Calculate the loss by Obtain a list of reference non-face features arranged in ascending order of loss. The reference non-face features at the top of the list have the strongest ability to represent masked non-face features, while the representative masked face features the weakest ability. In this way, the top M reference non-face features of the list can be continuously added to a feature pool P _- in an iterative manner to construct the final Preferably, M is greater than or equal to 1 and less than or equal to 50. Specifically, let the initial feature pool be empty or Then at step t use To select the top M candidates, get then, The features in are used to update Then used to solve the objective function in formula (5).

5) Merge dictionaries to obtain approximate external feature space

Among the above steps, steps 1) and 2) are not in strict order and can be performed sequentially or in parallel; steps 3) and 4) are not in strict order and can be performed successively or in parallel. Through the above steps, an approximate external feature space is constructed The approximate external feature space is composed of selecting the most representative features from a large number of reference face features and reference non-face features. The selection strategy is obtained by comparing with a large number of masked face features and masked non-face features. , the included features can well represent masked face features and can also distinguish masked non-face features, so using the approximate external feature space The embedded features obtained by embedding projection of candidate features have good representation ability for masked faces. On the other hand, compared with the traditional LLE method, the approximate external feature space constructed by the fast approximate LLE method proposed by the present invention is larger than the local feature space D _i corresponding to each candidate feature x _i , for each candidate feature x _i , after projective transformation, the obtained approximate embedded features Compared with the traditional LLE method, the dimension of the traditional embedding feature v _i is higher, and to a certain extent, it compensates for the loss of feature representation caused by the fast approximation. Therefore, the approximate external feature space constructed by the fast approximate LLE method proposed by the present invention It is used in masked face detection and has almost no impact on detection accuracy.

By comparing examples of representative reference face images and representative reference non-face images corresponding to the approximate external feature space, it can be found that the selected representative reference face images contain different appearance, wearing, skin color, Expressions, etc., so it can well represent masked faces and at the same time distinguish masked non-faces well; the selected representative reference non-face images are textured areas, incomplete faces, and more backgrounds Therefore, it can well represent masked non-human faces and distinguish masked faces well at the same time.

The above implementation is only used to illustrate the technical solution of the present invention and not to limit it. Those skilled in the art can modify or equivalently replace the technical solution of the present invention without departing from the spirit and scope of the present invention. Protection of the present invention The scope should be defined by the claims.

Claims (7)

1. a kind of method for detecting human face, step include:

1) candidate face detection is carried out to image to be detected, obtains candidate face image；

2) candidate feature extraction is carried out to the candidate face image, obtains candidate feature；

3) insertion transformation is carried out to the candidate feature, obtains tradition insertion feature or approximate insertion feature；

4) it to the tradition insertion feature or approximate insertion feature, is verified by classifying with regression algorithm, obtains detection knot Fruit；

Wherein, in step 3), after the candidate feature carries out insertion transformation by a surface space built in advance, Obtain tradition insertion feature or approximate insertion feature；The surface space is that conventional external feature space or proximate exterior are special Insertion transformation described in sign space is locally linear embedding into method or quick approximation is locally linear embedding into method realization using traditional；It passes The method that is locally linear embedding into of system carries out insertion transformation to the candidate feature with noise using conventional external feature space, is passed System insertion feature；Quick approximation, which is locally linear embedding into, is embedded in the candidate feature with noise using proximate exterior feature space Transformation obtains approximate insertion feature；

The quick approximation is locally linear embedding into the building method of proximate exterior feature space in method, comprising the following steps:

A) candidate face detection is carried out to the reference face data set marked and candidate feature is extracted, judge that candidate feature belongs to These candidate features are stored in reference to face characteristics dictionary respectively and refer to non-face feature by face characteristic or non-face feature Dictionary；

B) candidate face detection is carried out to the masked face data set marked and candidate feature is extracted, judge that candidate feature belongs to These candidate features are stored in masked face characteristics dictionary and masked non-by masked face characteristic or masked non-face feature respectively Face characteristic dictionary；

C) from above-mentioned with reference to selecting representative to represent above-mentioned masked face characteristics dictionary in face characteristics dictionary With reference to face characteristics dictionary；

D) from above-mentioned with reference to selecting representative to represent above-mentioned masked non-face tagged word in non-face characteristics dictionary The non-face characteristics dictionary of the reference of allusion quotation；

E) merging is above-mentioned representative with reference to face characteristics dictionary and representative with reference to non-face characteristics dictionary, obtains To proximate exterior feature space.

2. the method as described in claim 1, which is characterized in that in step a), by calculating the corresponding candidate of the candidate feature Degree of overlapping between face location and the face location marked determines that degree of overlapping is handed over and ratio is to measure, wherein friendship is simultaneously Candidate feature is judged for the feature with reference to face than being greater than 0.7, is handed over and than judging candidate feature for reference to inhuman less than 0.3 The feature of face.

3. the method as described in claim 1, which is characterized in that step b), by calculating the corresponding candidate of the candidate feature Degree of overlapping between face position and the face location marked determines that degree of overlapping is handed over and ratio is to measure, wherein hand over and compare Candidate feature is judged greater than 0.6 for the feature of masked face, handed over and than judging candidate feature to be masked non-face less than 0.4 Feature.

4. the method as described in claim 1, which is characterized in that use greedy algorithm from reference face characteristics dictionary in step c) The representative reference face characteristics dictionary of middle selection；The greedy algorithm refers to calculating with reference to each in face characteristics dictionary With reference to the loss of face characteristic, obtain by the reference face feature list for losing ascending ascending order arrangement, before taking the list most The reference face characteristic in face represents masked face characteristic；Wherein the loss refers to each reference face characteristic and masked face The distance of the arest neighbors feature of characteristics dictionary and each arest neighbors feature with reference to face characteristic and masked non-face characteristics dictionary Distance difference.

5. the method as described in claim 1, which is characterized in that using greedy algorithm from reference to non-face tagged word in step d) It is selected in allusion quotation representative with reference to non-face characteristics dictionary；The greedy algorithm refers to that calculating refers to non-face characteristics dictionary In each loss with reference to non-face feature, obtain taking by the non-face feature list of reference for losing the arrangement of ascending ascending order The non-face feature of the reference of the list foremost represents masked non-face feature；Wherein the loss refers to each with reference to inhuman Face feature is at a distance from the arest neighbors feature of masked non-face characteristics dictionary and each special with reference to non-face feature and masked face Levy the difference of the distance of the arest neighbors feature of dictionary.

6. a kind of human face detection device, including candidate block, insertion module and authentication module；

The candidate block is used to carry out candidate face detection to image to be detected and extracts candidate feature；

The insertion module obtains tradition insertion feature or approximate insertion is special for carrying out insertion transformation to the candidate feature Sign；

The authentication module is used to test above-mentioned tradition insertion feature or approximate insertion feature with regression algorithm by classifying Card, to obtain testing result to the end；

Wherein, the insertion module carries out insertion change to the candidate feature by a surface space built in advance After changing, tradition insertion feature or approximate insertion feature are obtained；The surface space is conventional external feature space or approximation Insertion transformation described in surface space is locally linear embedding into method or quick approximation is locally linear embedding into method using traditional It realizes；Traditional method that is locally linear embedding into carries out insertion change to the candidate feature with noise using conventional external feature space It changes, obtains tradition insertion feature；Quick approximation is locally linear embedding into special to the candidate with noise using proximate exterior feature space Sign carries out insertion transformation, obtains approximate insertion feature；

The quick approximation is locally linear embedding into the building method of proximate exterior feature space in method, comprising the following steps:

A) candidate face detection is carried out to the reference face data set marked and candidate feature is extracted, judge that candidate feature belongs to These candidate features are stored in reference to face characteristics dictionary respectively and refer to non-face feature by face characteristic or non-face feature Dictionary；

B) candidate face detection is carried out to the masked face data set marked and candidate feature is extracted, judge that candidate feature belongs to These candidate features are stored in masked face characteristics dictionary and masked non-by masked face characteristic or masked non-face feature respectively Face characteristic dictionary；

C) from above-mentioned with reference to selecting representative to represent above-mentioned masked face characteristics dictionary in face characteristics dictionary With reference to face characteristics dictionary；

D) from above-mentioned with reference to selecting representative to represent above-mentioned masked non-face tagged word in non-face characteristics dictionary The non-face characteristics dictionary of the reference of allusion quotation；

E) merging is above-mentioned representative with reference to face characteristics dictionary and representative with reference to non-face characteristics dictionary, obtains To proximate exterior feature space.

7. device as claimed in claim 6, which is characterized in that the candidate block obtains multiple candidate features, then embedding Enter after carrying out insertion transformation by a surface space built in advance in module, obtains tradition insertion feature or approximation It is embedded in feature；The surface space is conventional external feature space or proximate exterior feature space；The insertion transformation is adopted Method is locally linear embedding into or quick approximation is locally linear embedding into method and realizes with traditional.