CN109285149A

CN109285149A - Appraisal procedure, device and the calculating equipment of quality of human face image

Info

Publication number: CN109285149A
Application number: CN201811025831.9A
Authority: CN
Inventors: 郑俊君
Original assignee: Hangzhou Bi Zhi Technology Co Ltd
Current assignee: Hangzhou Bi Zhi Technology Co Ltd
Priority date: 2018-09-04
Filing date: 2018-09-04
Publication date: 2019-01-29

Abstract

The invention discloses a kind of appraisal procedure of quality of human face image, device and calculate equipment, comprising: obtain facial image to be assessed；The facial image is input to the quality testing model that training obtains, obtains at least one quality measurements of the multiple detection dimensions of correspondence of the quality testing model output；Wherein, the quality testing model includes the network for extracting feature and the network for detecting quality；Whether the quality for assessing the facial image according at least one described quality measurements is qualified.The present invention program can carry out prediction of quality from multiple detection dimensions using quality testing model, rather than be only capable of being predicted from single detection dimensions, so that the assessment result of face quality is more accurate；And directly facial image is input in quality testing model and carries out prediction of quality, the processing of the feature without extracting facial image in advance improves forecasting efficiency.

Description

Appraisal procedure, device and the calculating equipment of quality of human face image

Technical field

The present invention relates to field of computer technology, and in particular to a kind of appraisal procedure of quality of human face image, device and meter Calculate equipment.

Background technique

With the fast development of artificial intelligence (Artificial Intelligence, abbreviation AI), face recognition technology quilt It is applied to the every field such as mobile payment, member management, video monitoring, police criminal detection.But due to hardware limitation or the external world The influence of factor, cause the facial image of acquisition to there are following one or more problems: face is fuzzy, illumination is too strong/excessively weak, deposits Block, posture is excessive and resolution ratio is too low, and then influence recognition of face accuracy rate.As it can be seen that promoting the matter for comparing sample Amount promotes the quality of facial image, be the key factor for improving face recognition accuracy rate.

Currently, the detection of face quality mainly removes the fuzzy of detection face by the algorithm of some traditional image procossings Degree, the problems such as illumination is too strong or posture is excessive.The algorithm of these image procossings includes statistics with histogram, discrete cosine transform (Discrete Cosine Transform, abbreviation DCT), Sobel filtering, Gabor filtering, active appearance models (Active Appearance Model, abbreviation AAM), active shape model (Active Shape Model, abbreviation ASM) scheduling algorithm.

However, it is generally the case that these traditional image processing algorithms be only capable of for face qualitative factor in a certain respect into Row detection, such as AAM or ASM are only applicable to facial angle calculating, and poor for the detection accuracy of the face of wide-angle；For another example DCT only has application in fuzzy detection, and can not be applied to other face quality testings.

Summary of the invention

In view of the above problems, it proposes on the present invention overcomes the above problem or at least be partially solved in order to provide one kind State appraisal procedure, device and the calculating equipment of the quality of human face image of problem.

According to an aspect of the invention, there is provided a kind of appraisal procedure of quality of human face image, comprising:

Obtain facial image to be assessed；

The facial image is input to the quality testing model that training obtains, obtains the quality testing model output At least one quality measurements of corresponding multiple detection dimensions；Wherein, the quality testing model includes for extracting feature Network and network for detecting quality；

Whether the quality for assessing the facial image according at least one described quality measurements is qualified.

Further, the multiple detection dimensions include at least two in following multiple dimensions: fog-level dimension, light According to intensity dimension, posture dimension and sheltering part dimension.

Further, the quality testing model includes: that fisrt feature extracts network, corresponds to the multiple of multiple detection dimensions Second feature extracts multiple quality testing networks of network and corresponding multiple detection dimensions；

Wherein, each second feature is extracted network and is extracted from the face high-dimensional feature that fisrt feature extracts network output The face characteristic of corresponding detection dimensions.

Further, the quality testing model is obtained by following steps training:

Obtain multiple face sample images；

For every face sample image, the quality information of multiple detection dimensions of the face sample image is marked Note, obtains the quality annotation result of multiple detection dimensions of face sample image；

Every face sample image is input to fisrt feature to extract in network, obtains multiple quality testing network outputs The quality measurements of multiple detection dimensions of face sample image；

According to the multiple of the quality measurements of multiple detection dimensions of the face sample image and face sample image Loss between the quality annotation result of detection dimensions, to fisrt feature extract network, multiple second feature extract network and Multiple quality testing networks are trained, and obtain the quality testing model.

Further, the quality testing model includes: that fisrt feature extracts network, feature divides network and corresponding Multiple quality testing networks of multiple detection dimensions；

Wherein, the face characteristic that multiple quality testing networks divide the correspondence detection dimensions that network marks off to feature carries out Quality testing.

Further, the quality testing model is obtained by following steps training:

Obtain multiple face sample images；

According to the multiple of the quality measurements of multiple detection dimensions of the face sample image and face sample image Loss between the quality annotation result of detection dimensions extracts network to fisrt feature, feature divides network and multiple quality Detection network is trained, and obtains the quality testing model.

Further, described that the facial image is input to the quality testing model that training obtains, obtain the quality At least one quality measurements of the multiple detection dimensions of correspondence of detection model output specifically:

The facial image is input to fisrt feature and extracts network, obtains multiple matter of multiple quality testing network outputs Measure testing result；

Whether the quality that described at least one quality measurements according to assess the facial image is qualified specifically: Whether the quality for assessing the facial image according to multiple quality measurements that multiple quality testing networks export is qualified.

Further, the facial image is input to described before the quality testing model that training obtains, the side Method further include:

A current detection dimension is selected from multiple detection dimensions；Current detection dimension will be corresponded in quality testing model Quality testing network be determined as current detection network；

It is described that the facial image is input to the quality testing model that training obtains, it is defeated to obtain the quality testing model At least one quality measurements of the multiple detection dimensions of correspondence out: the facial image is input to fisrt feature and extracts net Network obtains a quality measurements of current Quality detection network output；

Whether the quality that described at least one quality measurements according to assess the facial image is qualified specifically: Judge whether a quality measurements of the current detection network output meet the quality standard of current detection dimension；

If a quality measurements of the current detection network output do not meet the quality standard of current detection dimension, Then the facial image is off quality.

Further, described that a current detection dimension is selected from multiple detection dimensions specifically: according to preset more The detection priority of a detection dimensions selects a non-selected current detection dimension from multiple detection dimensions；

The method also includes: if a quality measurements of current detection network output meet current detection dimension The quality standard of degree then jumps execution and is selected from multiple detection dimensions according to the detection priority of preset multiple detection dimensions The step of one non-selected current detection dimension and its subsequent step；Until judging the quality testing of multiple detection dimensions As a result meet the quality standard of corresponding detection dimensions, it is determined that the facial image it is up-to-standard.

According to another aspect of the present invention, a kind of assessment device of quality of human face image is provided, comprising:

Module is obtained, suitable for obtaining facial image to be assessed；

Prediction module obtains the quality suitable for the facial image is input to the quality testing model that training obtains At least one quality measurements of the multiple detection dimensions of correspondence of detection model output；Wherein, the quality testing model packet Include the network for extracting feature and the network for detecting quality；

Whether evaluation module, the quality suitable for assessing the facial image according at least one described quality measurements close Lattice.

Further, described device further include: training module is suitable for obtaining multiple face sample images；

Further, the prediction module is further adapted for:

The evaluation module is further adapted for: being assessed according to multiple quality measurements of multiple quality testing networks output Whether the quality of the facial image is qualified.

Further, described device further include:

Selecting module is suitable for selecting a current detection dimension from multiple detection dimensions；

Determining module, suitable for the quality testing network for corresponding to current detection dimension in quality testing model to be determined as currently Detect network；

The prediction module is further adapted for: the facial image being input to fisrt feature and extracts network, is obtained current One quality measurements of quality testing network output；

The evaluation module is further adapted for: judge current detection network output a quality measurements whether Meet the quality standard of current detection dimension；

Further, the selecting module is further adapted for: according to the detection priority of preset multiple detection dimensions from A non-selected current detection dimension is selected in multiple detection dimensions；

If a quality measurements of the current detection network output meet the quality standard of current detection dimension, The operation of triggering selection module.

According to another aspect of the invention, provide a kind of calculating equipment, comprising: processor, memory, communication interface and Communication bus, the processor, the memory and the communication interface complete mutual communication by the communication bus；

For the memory for storing an at least executable instruction, it is above-mentioned that the executable instruction executes the processor The corresponding operation of the appraisal procedure of quality of human face image.

In accordance with a further aspect of the present invention, provide a kind of computer storage medium, be stored in the storage medium to A few executable instruction, the executable instruction make processor execute the corresponding behaviour of appraisal procedure such as above-mentioned quality of human face image Make.

Appraisal procedure, device and the calculating equipment of quality of human face image according to the present invention, comprising: obtain people to be assessed Face image；The facial image is input to the quality testing model that training obtains, obtains the quality testing model output At least one quality measurements of corresponding multiple detection dimensions；Wherein, the quality testing model includes for extracting feature Network and network for detecting quality；The matter of the facial image is assessed according at least one described quality measurements Whether amount is qualified.The present invention program can carry out prediction of quality from multiple detection dimensions using quality testing model, rather than only It can be predicted from single detection dimensions, so that the assessment result of face quality is more accurate；And directly by face Image, which is input in quality testing model, carries out prediction of quality, the processing of the feature without extracting facial image in advance, Improve the efficiency of assessment.

The above description is only an overview of the technical scheme of the present invention, in order to better understand the technical means of the present invention, And it can be implemented in accordance with the contents of the specification, and in order to allow above and other objects of the present invention, feature and advantage can It is clearer and more comprehensible, the followings are specific embodiments of the present invention.

Detailed description of the invention

By reading the following detailed description of the preferred embodiment, various other advantages and benefits are common for this field Technical staff will become clear.The drawings are only for the purpose of illustrating a preferred embodiment, and is not considered as to the present invention Limitation.And throughout the drawings, the same reference numbers will be used to refer to the same parts.In the accompanying drawings:

Fig. 1 shows the flow chart of the appraisal procedure of quality of human face image according to an embodiment of the invention；

Fig. 2 shows the flow charts of the appraisal procedure of quality of human face image in accordance with another embodiment of the present invention；

Fig. 3 shows the schematic diagram of the training process of quality testing model in a specific embodiment of the invention；

Fig. 4 shows the flow chart of the appraisal procedure of the quality of human face image of another embodiment according to the present invention；

Fig. 5 shows the schematic diagram of the training process of quality testing model in another specific embodiment of the invention；

Fig. 6 shows the functional block diagram of the assessment device of quality of human face image according to an embodiment of the invention；

Fig. 7 shows a kind of structural schematic diagram for calculating equipment according to an embodiment of the present invention.

Specific embodiment

Exemplary embodiments of the present disclosure are described in more detail below with reference to accompanying drawings.Although showing the disclosure in attached drawing Exemplary embodiment, it being understood, however, that may be realized in various forms the disclosure without should be by embodiments set forth here It is limited.On the contrary, these embodiments are provided to facilitate a more thoroughly understanding of the present invention, and can be by the scope of the present disclosure It is fully disclosed to those skilled in the art.

Fig. 1 shows the flow chart of the appraisal procedure of quality of human face image according to an embodiment of the invention.Such as Fig. 1 institute Show, this method comprises:

Step S101 obtains facial image to be assessed.

Wherein, facial image to be assessed can be the acquisition of any imaging device, the face figure for needing to carry out quality evaluation Picture.Also, facial image to be assessed can be the original image of imaging device acquisition, or to carry out face to original image After the processing such as detection, what is obtained only includes the image of face, and the present invention is not specifically limited in this embodiment.

Facial image is input to the quality testing model that training obtains by step S102, obtains the output of quality testing model The multiple detection dimensions of correspondence at least one quality measurements.

Wherein, quality testing model includes the network for extracting feature and the network for detecting quality.And it is more A detection dimensions include influencing the corresponding dimension of image quality factors of face recognition accuracy, for example, the fog-level of image The accuracy for influencing recognition of face, then be determined as a detection dimensions for fog-level.

Specifically, it obtains including the network for extracting feature and the network for detecting quality by self study training Quality testing model in other words facial image can be directly inputted in the quality testing model, then by the net of extraction feature Network extracts the feature of the facial image of multiple detection dimensions, and is tieed up by the network of detection quality for multiple detections are extracted The feature of the facial image of at least one detection dimensions in degree is predicted, and exports at least one quality inspection that prediction obtains Survey result.

Quality testing model in the present invention can carry out prediction of quality for multiple detection dimensions, but the present invention is not The concrete mode of prediction is limited, in the specific implementation, those skilled in the art can be according to actual evaluation requirement and to assessment Accuracy requirement, determination is to carry out prediction of quality for multiple detection dimensions, or only for wherein preset quantity Detection dimensions carry out prediction of quality；Alternatively, determining for multiple detection dimensions is to carry out parallel anticipation, also according to preset suitable Sequence is serially predicted.

Whether step S103, the quality for assessing facial image according at least one quality measurements are qualified.

Specifically, according at least one quality measurements, and according to preset assessment rule, the matter of facial image is assessed Amount.The present invention does not do specific restriction to the rule of assessment, when it is implemented, those skilled in the art can be according to the reality of assessment Demand is flexibly set.Optionally, preset assessment rule includes corresponding according to meeting at least one quality measurements The quantity of the quality measurements of the quality standard of detection dimensions determines whether the quality of facial image is qualified.For example, preset Assessment rule is that the quality measurements of multiple detection dimensions meet the quality standard of corresponding detection dimensions, just can determine that face Image it is up-to-standard.

According to the appraisal procedure of quality of human face image provided in this embodiment, facial image to be assessed is obtained；By face Image is input to the obtained quality testing model of training, which includes network for extracting feature and be used for The network of quality, and then the processing of the feature without extracting facial image in advance are detected, and can be directly by facial image It inputs and carries out prediction of quality in the model, simplify the algorithm flow of quality evaluation, improve the efficiency estimated；And it utilizes Quality testing model can carry out prediction of quality from multiple detection dimensions, rather than be only capable of carrying out from single detection dimensions pre- It surveys, so that the assessment result of face quality is more accurate, is more advantageous to the accuracy for improving recognition of face.

Fig. 2 shows the flow charts of the appraisal procedure of quality of human face image in accordance with another embodiment of the present invention.This reality It applies in example, the feature of multiple detection dimensions extracts network by multiple second feature in quality testing model and extracts to obtain. As shown in Fig. 2, this method comprises:

Step S201, training obtain quality testing model；Wherein, quality testing model include: fisrt feature extract network, Multiple second feature of corresponding multiple detection dimensions extract multiple quality testing nets of network and corresponding multiple detection dimensions Network.

Wherein, it is that shared face high-dimensional feature extracts network that fisrt feature, which extracts network, and each second feature is extracted Network extracts the face characteristic that corresponding detection dimensions are extracted in the face high-dimensional feature that network exports from fisrt feature.Also, In the present embodiment, it can be any convolutional neural networks that fisrt feature, which extracts network and/or each second feature extraction network, Structure, optionally, fisrt feature extract network and/or each second feature extract network can be ResNet, DenseNet, The common neural network such as Vgg or RetinaNet, but the present invention is not limited thereto.And corresponding multiple detection dimensions Multiple quality testing networks in prediction mode can be the mode of classification, be also possible to the mode returned or two kinds Mode one is reinstated, and the present invention is also not specifically limited this.

Wherein, multiple detection dimensions include at least two in following multiple dimensions: fog-level dimension, intensity of illumination dimension Degree, posture dimension and sheltering part dimension.

Specifically, quality testing model is obtained by shared backbone pattern drill: step 1 obtains multiple face samples This image.Further, face sample image is obtained from the facial image that a variety of different imaging devices acquire, so that training Obtained quality testing model can be used for carrying out prediction of quality to the facial image that various imaging devices acquire, and then improve matter Measure the scope of application of the quality testing of detection model.Step 2, for every face sample image, to the more of face sample image The quality information of a detection dimensions is labeled, and obtains the quality annotation result of multiple detection dimensions of face sample image.Into One step, the quality information of each detection dimensions is labeled, optionally, fog-level, illumination to face sample image Intensity, angular deflection situation are (including the angle around X-axis rotation, around the angle of Y-axis rotation, and/or around Z axis rotation Angle) and/or sheltering part be labeled.Every face sample image is input to fisrt feature and extracted in network by step 3, Obtain the quality measurements of multiple detection dimensions of the face sample image of multiple quality testing network outputs.Further, Face sample image is input to fisrt feature to extract in network, fisrt feature extracts network and extracts face high-dimensional feature；So Each second feature extracts network and extracts corresponding detection dimensions from the face high-dimensional feature that fisrt feature network exports afterwards Face characteristic finally extracts the face that network extracts according to the second feature of corresponding detection dimensions by each quality testing network Feature carries out prediction of quality.Step 4, according to the quality measurements of multiple detection dimensions of face sample image and face sample Loss between the quality annotation result of multiple detection dimensions of this image extracts network, multiple second feature to fisrt feature It extracts network and multiple quality testing networks is trained, obtain quality testing model.Further, according to the big of penalty values Small progress backpropagation and training optionally by the derivative of penalty values in a manner of backpropagation, adjust each matter from lower to upper Amount prediction network and each feature extraction network in parameter, such loop iteration is multiple, until penalty values meet it is preset It is required that and then stop network training, and save each network parameter generate quality testing model.Further, Duo Gezhi Multiple quality measurements of the corresponding multiple detection dimensions of amount detection network output, can be by by the quality of each detection dimensions Testing result and quality annotation result carry out difference calculating, and are lost according to the difference calculated result of multiple detection dimensions Value；Alternatively, assigning corresponding training weight for each detection dimensions, multiple quality annotation results and corresponding trained weight are calculated The sum of products, and calculate the sum of products of multiple quality measurements and corresponding trained weight, and knot will be calculated twice The difference of fruit is determined as penalty values.Wherein, the mode that above-mentioned penalty values calculate includes but is not limited to mean square error (mean- Square error, abbreviation MSE), cosine distance, Euclidean distance, L1 distance or L2 distance etc. modes.

For the understanding convenient for obtaining the process of quality testing model to above-mentioned training, specifically illustrated with one below The process.Fig. 3 shows the schematic diagram of the training process of quality testing model in a specific embodiment of the invention.Such as Fig. 3 institute Show, it is that shared face high-dimensional feature extracts network that fisrt feature, which extracts network, and fog-level dimensional characteristics extract network, light Network is extracted according to strength characteristic, posture dimensional characteristics extract network and sheltering part dimensional characteristics extract network respectively from first The feature of corresponding detection dimensions is extracted in the face high-dimensional feature that feature extraction network extracts；Fog-level dimension quality Detect network, intensity of illumination quality testing network, posture dimension quality testing network and sheltering part dimension quality testing net The face characteristic for the detection dimensions that network is extracted according to corresponding feature extraction network respectively carries out prediction of quality；Loss It calculates network and penalty values is calculated according to the quality measurements and annotation results of multiple detection dimensions, according to the penalty values It adjusts fisrt feature and extracts network, multiple second feature extraction network of corresponding multiple detection dimensions and corresponding multiple detections The parameter of multiple quality testing networks of dimension, until penalty values meet preset requirement, at this point, then saving fisrt feature extracts net Multiple quality testings of network, multiple second feature extraction network of corresponding multiple detection dimensions and corresponding multiple detection dimensions The parameter of network, and fisrt feature extracted into network, multiple second feature of corresponding multiple detection dimensions extract network and right Answer being determined entirely by as quality testing model of multiple quality testing networks of multiple detection dimensions.

Step S202 obtains facial image to be assessed.

After imaging device is imaged, human face region is detected using Face datection algorithm, will test Human face region image is as facial image to be assessed.Wherein, imaging device can be any equipment that can generate image, packet Include but be not limited to the imaging devices such as IP Camera, mobile phone camera, camera, near-infrared camera.

Facial image is input to the quality testing model that training obtains by step S203, obtains the output of quality testing model The multiple detection dimensions of correspondence at least one quality measurements.

Specifically, after obtaining quality testing model, facial image is input to the fisrt feature of quality testing model It extracts in network, and obtains the quality measurements exported by least one quality testing network of the multiple detection dimensions of correspondence. In the present invention, specific detection mode when being detected using the quality testing model is not limited, it optionally, can same hour hands (parallel form i.e. hereinafter) is detected to multiple detection dimensions, alternatively, can also be with preset detection ordering successively needle (serial manner i.e. hereinafter) is detected to each detection dimensions, when it is implemented, those skilled in the art can be flexible Select the mode of detection.

Further, it is detected according to parallel form, then facial image is input to fisrt feature and extracts network, Obtain multiple quality measurements of multiple quality testing network outputs.It is detected according to serial manner, then from multiple A current detection dimension is selected in detection dimensions；The quality testing network of current detection dimension will be corresponded in quality testing model It is determined as current detection network；Facial image is input to fisrt feature and extracts network, obtains current Quality detection network output A quality measurements.Optionally, in parallel form, after current detection network has been determined, by current detection The detection weight assignment of network is the first weight, by the detection weight assignment of the quality testing network in addition to current detection network For the second weight, for example, the first weight is 1, the second weight is 0, by the processing of the weight assignment, be may be implemented only for working as Preceding network carries out quality testing, obtains a quality measurements of current detection network output.Further, selection is worked as The process of preceding detection dimensions can select one according to the detection priority of preset multiple detection dimensions from multiple detection dimensions Non-selected current detection dimension is realized and is detected according to the detection ordering of detection priority from high to low.

Whether step S204, the quality for assessing facial image according at least one quality measurements are qualified.

Specifically, the quality measurements of at least one quality testing network output in multiple quality testing networks are pair The testing result for the detection dimensions answered, and when assessing the quality of facial image, then it is to be commented according to preset assessment rule Estimate, optionally, is assessed according to the quantity of the quality measurements for the quality standard for meeting corresponding detection dimensions, for example, having More than the quality standard that the quality measurements of the detection dimensions of preset ratio meet corresponding detection dimensions, then facial image is assessed It is up-to-standard；Or assessed according to the corresponding detection dimensions of the quality measurements met the quality standard, for example, detection The quality measurements of the detection dimensions of the preset quantity of highest priority are met the quality standard, then assess the quality of facial image It is qualified.

Further, defeated according to multiple quality testing networks if being detected in step S203 using parallel form Whether the quality of multiple quality measurements assessment facial image out is qualified.Multiple matter that multiple quality testing networks are exported Amount testing result is compared with the quality standard of corresponding detection dimensions, if being more than quality standard, it is determined that corresponding detection dimension That spends is up-to-standard, and then whether the quality according to the quantity of qualified detection dimensions assessment facial image is qualified.Alternatively, calculating The difference of the quality measurements of multiple detection dimensions and the quality standard of corresponding detection dimensions, and the assessment of corresponding detection dimensions Whether the sum of products of weight is qualified according to calculated result and the quality of the size assessment facial image of preset quality value.With Fig. 3 For obtained quality testing model, preset quality value is 3, when calculated result is less than the preset quality value, it is determined that face figure Picture it is up-to-standard, detected using parallel form, then simultaneously obtain four detection dimensions quality testing network output Quality measurements, it is right if four quality measurements and the difference of the quality standard of corresponding detection dimensions are 3,6,1,2 The assessment weight for four detection dimensions answered is 0.5,0.1,0.2,0.2, then the sum of products is 3*0.5+6*0.1+1*0.2+2* 0.2=2.7, then assessment result is the up-to-standard of facial image.As it can be seen that this parallel detection mode, can only pass through one It is secondary to detect to get assessment result is arrived, and assessed according to multiple testing results, the accuracy of assessment result can be improved.

Further, if being detected in step S203 using serial manner, judge the output of current detection network Whether one quality measurements meets the quality standard of current detection dimension；If the quality inspection of current detection network output The quality standard that result does not meet current detection dimension is surveyed, then facial image is off quality.Wherein, according to detection priority Height successively select current detection dimension, if the quality measurements of the high detection dimensions of the detection priority currently selected are not It meets the quality standard, then without carrying out the detection of the low detection dimensions of remaining detection priority, and then the fortune of detection can be reduced Calculation amount, while can quickly find underproof facial image.Further, if a quality of current detection network output Testing result meets the quality standard of current detection dimension, then it is preferential according to the detection of preset multiple detection dimensions to jump execution Grade selects the step of non-selected current detection dimension and its subsequent step from multiple detection dimensions；Until judging The quality measurements of multiple detection dimensions meet the quality standard of corresponding detection dimensions, it is determined that the quality of facial image is closed Lattice.Still by taking the quality testing model that Fig. 3 is obtained as an example, if fog-level dimension, intensity of illumination dimension, appearance from left to right in Fig. 3 The detection priority of state dimension and sheltering part dimension is reduced with this, then choosing fog-level detection dimensions first is current inspection Dimension is surveyed, determines that the fog-level dimension of facial image meets matter according to the quality measurements that fog-level detects network output Amount standard then selects intensity of illumination dimension for current detection dimension, and the quality testing knot of network output is detected according to intensity of illumination Fruit determines that the intensity of illumination dimension of facial image is met the quality standard ... until determining the last one detection dimension of facial image Degree (sheltering part dimension) also complies with quality standard, just can determine that the up-to-standard of facial image.

The present embodiment learns a kind of quality testing model by the method training of deep learning, obtains matter using the training The quality evaluation that detection model carries out facial image is measured, is had in algorithm accuracy rate compared to traditional image processing algorithm bright Aobvious promotion, and it is suitable for the situation of various complexity, there is very high robustness, and traditional algorithm must be in special scenes Lower use, and have using limitation.In addition, the quality testing model can the quality precisely to multiple detection dimensions examine It surveys, and then is able to carry out information feedback, to be applied to user guided or interaction.

According to the appraisal procedure of quality of human face image provided in this embodiment, quality testing model includes: that fisrt feature mentions Take network, multiple second feature of corresponding multiple detection dimensions extract network and correspond to multiple quality of multiple detection dimensions Detect network, wherein fisrt feature extracts network for extracting shared face high-dimensional feature；Also, matter is obtained in training During measuring detection model, fisrt feature is adjusted by way of backpropagation and extracts network, multiple second feature extraction net Parameter in network and multiple quality testing networks, so that the quality testing model that training obtains can not only carry out facial image Quality testing, and can directly against facial image carry out feature extraction, simplify the algorithm flow of quality evaluation, improve The efficiency estimated；And when carrying out quality testing using quality testing model, detection mode is flexible, can be according to detection Demand selection is detected using the mode of serial or parallel, wherein parallel detection mode can only pass through one-time detection, i.e., Assessment result is obtained, and is assessed according to multiple testing results, the accuracy of assessment result can be improved；Serial detection Mode, it is possible to reduce the operand of detection, while can quickly find underproof facial image.

Fig. 4 shows the flow chart of the appraisal procedure of the quality of human face image of another embodiment according to the present invention.This reality It applies in example, the feature of multiple detection dimensions divides network by the feature in quality testing model and divided to obtain.Such as Fig. 4 institute Show, this method comprises:

Step S401, training obtain quality testing model；Wherein, quality testing model include: fisrt feature extract network, Feature divides multiple quality testing networks of network and corresponding multiple detection dimensions.

Wherein, it is that shared face high-dimensional feature extracts network, multiple quality testing networks that fisrt feature, which extracts network, The face characteristic for dividing the correspondence detection dimensions that network marks off to feature carries out quality testing.And multiple detection dimensions packets Include at least two in following multiple dimensions: fog-level dimension, intensity of illumination dimension, posture dimension and sheltering part dimension Degree.

Specifically, the training process of quality testing model is substantially consistent with the training step in step S201, below only needle The difference of the two is described in detail, and identical process then repeats no more, for details, reference can be made to saying in step S201 It is bright.Step 1 obtains multiple face sample images.Step 2, for every face sample image, to the more of face sample image The quality information of a detection dimensions is labeled, and obtains the quality annotation result of multiple detection dimensions of face sample image.Step Rapid three, every face sample image is input to fisrt feature and is extracted in network, the people of multiple quality testing network outputs is obtained The quality measurements of multiple detection dimensions of face sample image.Further, face sample image is input to fisrt feature It extracts in network, fisrt feature extracts network and extracts face high-dimensional feature；Then feature divides network to fisrt feature network The face high-dimensional feature of output is divided, and the face characteristic of division is input to the quality testing net of corresponding detection dimensions In network；Prediction of quality is finally carried out according to the face characteristic of the correspondence detection dimensions marked off by each quality testing network.Step Rapid four, according to multiple detection dimensions of the quality measurements of multiple detection dimensions of face sample image and face sample image Quality annotation result between loss, network is extracted to fisrt feature, feature divides network and multiple quality testing networks It is trained, obtains quality testing model.Wherein, network is divided to the feature in quality detection model to be trained i.e. by not Disconnected self study, determines the degree of association of the quality of each feature and each detection dimensions, and is divided to feature according to the degree of association Corresponding detection dimensions, to carry out the quality testing of the detection dimensions.

For the understanding convenient for obtaining the process of quality testing model to above-mentioned training, specifically illustrated with one below The process.Fig. 5 shows the schematic diagram of the training process of quality testing model in another specific embodiment of the invention.Such as Fig. 5 Shown, it is that shared face high-dimensional feature extracts network that fisrt feature, which extracts network, and it is high-dimensional from face that feature divides network The face characteristic of corresponding multiple detection dimensions is marked off in feature；Fog-level dimension quality testing network, intensity of illumination quality Detection network, posture dimension quality testing network and sheltering part dimension quality testing network obtain accordingly according to division respectively The face characteristic of detection dimensions carries out prediction of quality；Costing bio disturbance network according to the quality measurements of multiple detection dimensions and Penalty values are calculated in annotation results, adjust fisrt feature extraction network according to the penalty values, feature divides network and corresponding The parameter of multiple quality testing networks of multiple detection dimensions, until penalty values meet preset requirement, at this point, it is special then to save first Sign extracts network, feature divides network and the parameter of multiple quality testing networks of corresponding multiple detection dimensions, and by first Feature extraction network, feature divide multiple quality testing networks of network and corresponding multiple detection dimensions be determined entirely by for Quality testing model.

It should be noted that the training process of Fig. 3 and Fig. 5 is exemplary only, in actual implementation, it can also be added Its detection dimensions is trained, such as, if it closes one's eyes, and then more fully quality testing can be carried out to facial image.

Step S402 obtains facial image to be assessed.

Facial image is input to the quality testing model that training obtains by step S403, obtains the output of quality testing model The multiple detection dimensions of correspondence at least one quality measurements.

Whether step S404, the quality for assessing facial image according at least one quality measurements are qualified.

The concrete principle and process of above-mentioned steps S402 to step S404 is identical to step S204 as step S202, specifically may be used Referring to step S202 to the description of step S204, details are not described herein again.

According to the appraisal procedure of quality of human face image provided in this embodiment, embodiment corresponding with Fig. 2 the difference is that In the present embodiment, quality testing model includes: that fisrt feature extracts network, feature divides network and corresponding multiple detections Multiple quality testing networks of dimension, wherein it is that shared face high-dimensional feature extracts network that fisrt feature, which extracts network, more The face characteristic that a quality testing network divides the correspondence detection dimensions that network marks off to feature carries out quality testing；Also, Feature divides network and is also obtained by training, and then can accurately obtain each feature and multiple detections according to big data when training The degree of association of the quality of dimension can then carry out face high-dimensional feature according to detection dimensions quasi- when carrying out quality testing It really divides, is more advantageous to the accuracy for improving testing result.

In fig. 2 above and the corresponding embodiment of Fig. 4, only carried out using shared fisrt feature extraction module as example Illustrate, but in the specific implementation, independent high-dimensional feature can also be possessed with each detection dimensions and extracts network.

Fig. 6 shows the functional block diagram of the assessment device of quality of human face image according to an embodiment of the invention.Such as Fig. 6 Shown, which includes:

Module 601 is obtained, suitable for obtaining facial image to be assessed；

Prediction module 602 obtains the matter suitable for the facial image is input to the quality testing model that training obtains Measure at least one quality measurements of the multiple detection dimensions of correspondence of detection model output；Wherein, the quality testing model Including the network for extracting feature and the network for detecting quality；

Evaluation module 603, the quality suitable for assessing the facial image according at least one described quality measurements are No qualification.

In a kind of optional embodiment, the multiple detection dimensions include at least two in following multiple dimensions: Fog-level dimension, intensity of illumination dimension, posture dimension and sheltering part dimension.

In a kind of optional embodiment, the quality testing model include: fisrt feature extract network, correspondence it is multiple Multiple second feature of detection dimensions extract multiple quality testing networks of network and corresponding multiple detection dimensions；

In a kind of optional embodiment, described device further include: training module 604 is suitable for obtaining multiple face samples This image；

In a kind of optional embodiment, the quality testing model includes: that fisrt feature extracts network, feature divides Multiple quality testing networks of network and corresponding multiple detection dimensions；

In a kind of optional embodiment, the prediction module 602 is further adapted for:

The evaluation module 603 is further adapted for: the multiple quality measurements exported according to multiple quality testing networks Whether the quality for assessing the facial image is qualified.

In a kind of optional embodiment, described device further include:

Selecting module 605 is suitable for selecting a current detection dimension from multiple detection dimensions；

Determining module 606, suitable for the quality testing network for corresponding to current detection dimension in quality testing model to be determined as Current detection network；

The prediction module 602 is further adapted for: the facial image being input to fisrt feature and extracts network, is worked as One quality measurements of preceding quality testing network output；

The evaluation module 603 is further adapted for: judging a quality measurements of the current detection network output Whether the quality standard of current detection dimension is met；

In a kind of optional embodiment, the selecting module 605 is further adapted for: being tieed up according to preset multiple detections The detection priority of degree selects a non-selected current detection dimension from multiple detection dimensions；

It can refer to the description of corresponding steps in embodiment of the method about the specific structure and working principle of above-mentioned modules, Details are not described herein again.

The embodiment of the present application provides a kind of nonvolatile computer storage media, and the computer storage medium is stored with The quality of human face image in above-mentioned any means embodiment can be performed in an at least executable instruction, the computer executable instructions Appraisal procedure.

Fig. 7 shows a kind of structural schematic diagram for calculating equipment according to an embodiment of the present invention, the specific embodiment of the invention The specific implementation for calculating equipment is not limited.

As shown in fig. 7, the calculating equipment may include: processor (processor) 702, communication interface (Communications Interface) 704, memory (memory) 706 and communication bus 708.

Wherein:

Processor 702, communication interface 704 and memory 706 complete mutual communication by communication bus 708.

Communication interface 704, for being communicated with the network element of other equipment such as client or other servers etc..

Processor 702, for executing program 710, the appraisal procedure that can specifically execute above-mentioned quality of human face image is implemented Correlation step in example.

Specifically, program 710 may include program code, which includes computer operation instruction.

Processor 702 may be central processor CPU or specific integrated circuit ASIC (Application Specific Integrated Circuit), or be arranged to implement the integrated electricity of one or more of the embodiment of the present invention Road.The one or more processors that equipment includes are calculated, can be same type of processor, such as one or more CPU；It can also To be different types of processor, such as one or more CPU and one or more ASIC.

Memory 706, for storing program 710.Memory 706 may include high speed RAM memory, it is also possible to further include Nonvolatile memory (non-volatile memory), for example, at least a magnetic disk storage.

Program 710 specifically can be used for so that processor 702 executes following operation:

Obtain facial image to be assessed；

In a kind of optional embodiment, wherein the multiple detection dimensions include in following multiple dimensions at least Two: fog-level dimension, intensity of illumination dimension, posture dimension and sheltering part dimension.

In a kind of optional embodiment, wherein the quality testing model includes: that fisrt feature extracts network, right Multiple second feature of multiple detection dimensions are answered to extract multiple quality testing networks of network and corresponding multiple detection dimensions；

In a kind of optional embodiment, program 710 can specifically be further used for so that processor 702 execute it is following Operation: multiple face sample images are obtained；

In a kind of optional embodiment, program 710 can specifically be further used for so that processor 702 execute it is following Operation:

Obtain multiple face sample images；

According to multiple quality testing networks export multiple quality measurements assess the facial image quality whether It is qualified.

The facial image is input to fisrt feature and extracts network, obtains a matter of current Quality detection network output Measure testing result；

Judge whether a quality measurements of the current detection network output meet the quality of current detection dimension Standard；

In a kind of optional embodiment, program 710 can specifically be further used for so that processor 702 execute it is following Operation: select from multiple detection dimensions according to the detection priority of preset multiple detection dimensions one it is non-selected currently Detection dimensions；

If a quality measurements of the current detection network output meet the quality standard of current detection dimension, Jump execution selected from multiple detection dimensions according to the detection priority of preset multiple detection dimensions one it is non-selected The step of current detection dimension and its subsequent step；Until judging that the quality measurements of multiple detection dimensions meet correspondence The quality standard of detection dimensions, it is determined that the facial image it is up-to-standard.

Algorithm and display are not inherently related to any particular computer, virtual system, or other device provided herein. Various general-purpose systems can also be used together with teachings based herein.As described above, it constructs required by this kind of system Structure be obvious.In addition, the present invention is also not directed to any particular programming language.It should be understood that can use various Programming language realizes summary of the invention described herein, and the description done above to language-specific is to disclose this hair Bright preferred forms.

In the instructions provided here, numerous specific details are set forth.It is to be appreciated, however, that implementation of the invention Example can be practiced without these specific details.In some instances, well known method, structure is not been shown in detail And technology, so as not to obscure the understanding of this specification.

Similarly, it should be understood that in order to simplify the disclosure and help to understand one or more of the various inventive aspects, Above in the description of exemplary embodiment of the present invention, each feature of the invention is grouped together into single implementation sometimes In example, figure or descriptions thereof.However, the disclosed method should not be interpreted as reflecting the following intention: i.e. required to protect Shield the present invention claims features more more than feature expressly recited in each claim.More precisely, as following Claims reflect as, inventive aspect is all features less than single embodiment disclosed above.Therefore, Thus the claims for following specific embodiment are expressly incorporated in the specific embodiment, wherein each claim itself All as a separate embodiment of the present invention.

Those skilled in the art will understand that can be carried out adaptively to the module in the equipment in embodiment Change and they are arranged in one or more devices different from this embodiment.It can be the module or list in embodiment Member or component are combined into a module or unit or component, and furthermore they can be divided into multiple submodule or subelement or Sub-component.Other than such feature and/or at least some of process or unit exclude each other, it can use any Combination is to all features disclosed in this specification (including adjoint claim, abstract and attached drawing) and so disclosed All process or units of what method or apparatus are combined.Unless expressly stated otherwise, this specification is (including adjoint power Benefit require, abstract and attached drawing) disclosed in each feature can carry out generation with an alternative feature that provides the same, equivalent, or similar purpose It replaces.

In addition, it will be appreciated by those of skill in the art that although some embodiments described herein include other embodiments In included certain features rather than other feature, but the combination of the feature of different embodiments mean it is of the invention Within the scope of and form different embodiments.For example, in the following claims, embodiment claimed is appointed Meaning one of can in any combination mode come using.

Various component embodiments of the invention can be implemented in hardware, or to run on one or more processors Software module realize, or be implemented in a combination thereof.It will be understood by those of skill in the art that can be used in practice Microprocessor or digital signal processor (DSP) realize the assessment device of quality of human face image according to an embodiment of the present invention In some or all components some or all functions.The present invention is also implemented as described herein for executing Some or all device or device programs (for example, computer program and computer program product) of method.In this way Realization program of the invention can store on a computer-readable medium, or can have the shape of one or more signal Formula.Such signal can be downloaded from an internet website to obtain, and perhaps be provided on the carrier signal or with any other shape Formula provides.

It should be noted that the above-mentioned embodiments illustrate rather than limit the invention, and ability Field technique personnel can be designed alternative embodiment without departing from the scope of the appended claims.In the claims, Any reference symbol between parentheses should not be configured to limitations on claims.Word "comprising" does not exclude the presence of not Element or step listed in the claims.Word "a" or "an" located in front of the element does not exclude the presence of multiple such Element.The present invention can be by means of including the hardware of several different elements and being come by means of properly programmed computer real It is existing.In the unit claims listing several devices, several in these devices can be through the same hardware branch To embody.The use of word first, second, and third does not indicate any sequence.These words can be explained and be run after fame Claim.

Claims

1. a kind of appraisal procedure of quality of human face image, comprising:

Obtain facial image to be assessed；

The facial image is input to the quality testing model that training obtains, obtains the correspondence of the quality testing model output At least one quality measurements of multiple detection dimensions；Wherein, the quality testing model includes the net for extracting feature Network and network for detecting quality；

2. according to the method described in claim 1, wherein, the multiple detection dimensions include at least two in following multiple dimensions It is a: fog-level dimension, intensity of illumination dimension, posture dimension and sheltering part dimension.

3. method according to claim 1 or 2, wherein the quality testing model include: fisrt feature extract network, Multiple second feature of corresponding multiple detection dimensions extract multiple quality testing nets of network and corresponding multiple detection dimensions Network；

Wherein, each second feature extracts network and extracts extraction correspondence in the face high-dimensional feature that network exports from fisrt feature The face characteristic of detection dimensions.

4. according to the method described in claim 3, wherein, the quality testing model is obtained by following steps training:

Obtain multiple face sample images；

For every face sample image, the quality information of multiple detection dimensions of the face sample image is labeled, Obtain the quality annotation result of multiple detection dimensions of face sample image；

Every face sample image is input to fisrt feature to extract in network, obtains the face of multiple quality testing network outputs The quality measurements of multiple detection dimensions of sample image；

According to multiple detections of the quality measurements of multiple detection dimensions of the face sample image and face sample image Loss between the quality annotation result of dimension extracts network to fisrt feature, multiple second feature extract network and multiple Quality testing network is trained, and obtains the quality testing model.

5. method according to claim 1 or 2, wherein the quality testing model include: fisrt feature extract network, Feature divides multiple quality testing networks of network and corresponding multiple detection dimensions；

Wherein, the face characteristic that multiple quality testing networks divide the correspondence detection dimensions that network marks off to feature carries out quality Detection.

6. according to the method described in claim 5, wherein, the quality testing model is obtained by following steps training:

Obtain multiple face sample images；

According to multiple detections of the quality measurements of multiple detection dimensions of the face sample image and face sample image Loss between the quality annotation result of dimension extracts network to fisrt feature, feature divides network and multiple quality testings Network is trained, and obtains the quality testing model.

7. the method according to claim 3 or 5, wherein described that the facial image is input to the quality that training obtains Detection model, at least one quality measurements for obtaining the multiple detection dimensions of correspondence of the quality testing model output are specific Are as follows:

The facial image is input to fisrt feature and extracts network, obtains multiple quality inspection of multiple quality testing network outputs Survey result；

Whether the quality that described at least one quality measurements according to assess the facial image is qualified specifically: according to Whether the quality that multiple quality measurements of multiple quality testing network outputs assess the facial image is qualified.

8. the method according to claim 3 or 5, wherein the facial image is input to the matter that training obtains described Before measuring detection model, the method also includes:

A current detection dimension is selected from multiple detection dimensions；The matter of current detection dimension will be corresponded in quality testing model Amount detection network is determined as current detection network；

It is described that the facial image is input to the quality testing model that training obtains, obtain the quality testing model output At least one quality measurements of corresponding multiple detection dimensions: being input to fisrt feature for the facial image and extract network, Obtain a quality measurements of current Quality detection network output；

Whether the quality that described at least one quality measurements according to assess the facial image is qualified specifically: judgement Whether one quality measurements of the current detection network output meet the quality standard of current detection dimension；

If a quality measurements of the current detection network output do not meet the quality standard of current detection dimension, institute State the off quality of facial image.

9. according to the method described in claim 8, wherein, described one current detection dimension of selection from multiple detection dimensions has Body are as follows: selected from multiple detection dimensions according to the detection priority of preset multiple detection dimensions one it is non-selected current Detection dimensions；

The method also includes: if a quality measurements of current detection network output meet current detection dimension Quality standard then jumps execution according to the detection priority of preset multiple detection dimensions and selects one from multiple detection dimensions The step of non-selected current detection dimension and its subsequent step；Until judging the quality measurements of multiple detection dimensions Meet the quality standard of corresponding detection dimensions, it is determined that the facial image it is up-to-standard.

10. a kind of assessment device of quality of human face image, comprising:

Module is obtained, suitable for obtaining facial image to be assessed；

Prediction module obtains the quality testing suitable for the facial image is input to the quality testing model that training obtains At least one quality measurements of the multiple detection dimensions of correspondence of model output；Wherein, the quality testing model includes using In the network and network for detecting quality that extract feature；

Whether evaluation module, the quality suitable for assessing the facial image according at least one described quality measurements are qualified.

11. device according to claim 10, wherein the multiple detection dimensions include in following multiple dimensions at least Two: fog-level dimension, intensity of illumination dimension, posture dimension and sheltering part dimension.

12. device described in 0 or 11 according to claim 1, wherein the quality testing model includes: that fisrt feature extracts net Multiple quality testings of network, multiple second feature extraction network of corresponding multiple detection dimensions and corresponding multiple detection dimensions Network；

13. device according to claim 12, wherein described device further include: training module is suitable for obtaining multiple faces Sample image；

14. device described in 0 or 11 according to claim 1, wherein the quality testing model includes: that fisrt feature extracts net Network, feature divide multiple quality testing networks of network and corresponding multiple detection dimensions；

15. device according to claim 14, wherein described device further include: training module is suitable for obtaining multiple faces Sample image；

16. device described in 2 or 14 according to claim 1, wherein the prediction module is further adapted for:

The evaluation module is further adapted for: according to multiple quality measurements assessment that multiple quality testing networks export Whether the quality of facial image is qualified.

17. device described in 2 or 14 according to claim 1, wherein described device further include:

Determining module, suitable for the quality testing network for corresponding to current detection dimension in quality testing model is determined as current detection Network；

The prediction module is further adapted for: the facial image being input to fisrt feature and extracts network, obtains current Quality Detect a quality measurements of network output；

The evaluation module is further adapted for: judging whether a quality measurements of the current detection network output meet The quality standard of current detection dimension；

18. device according to claim 17, wherein the selecting module is further adapted for: according to preset multiple inspections The detection priority for surveying dimension selects a non-selected current detection dimension from multiple detection dimensions；

If a quality measurements of the current detection network output meet the quality standard of current detection dimension, trigger Selecting module operation.

19. a kind of calculating equipment, comprising: processor, memory, communication interface and communication bus, the processor, the storage Device and the communication interface complete mutual communication by the communication bus；

The memory executes the processor as right is wanted for storing an at least executable instruction, the executable instruction Ask the corresponding operation of the appraisal procedure of quality of human face image described in any one of 1-9.

20. a kind of computer storage medium, an at least executable instruction, the executable instruction are stored in the storage medium Processor is set to execute the corresponding operation of appraisal procedure of quality of human face image as claimed in any one of claims 1-9 wherein.