CN106384080A - Apparent age estimating method and device based on convolutional neural network - Google Patents

Abstract

The invention discloses an apparent age estimating method based on the convolutional neural network. The method comprises the steps that a) face detection is carried out, and the size of a face picture is determined; b) an image is preprocessed, and the mean RGB value of each pixel is calculated in a training set; c) the mean RGB value of each pixel is input to a multi-convolution layer; d) pooling is carried out in a pooling layer; e) a full connection layer is entered, and two adjacent layers are connected fully via a weight; f) a softmax loss function is used for monitoring a structure deeply; g) time-delay reverse conduction is used to reduce information loss; h) a hinge loss function is used to monitor regression and further reduce information loss; and i) a probability matrix is used. Thus, the softmax loss function is used for monitoring the structure deeply; a time-delay conduction mechanism reduces conduction layers to reduce information loss; a high loss module reduces conduction layers to reduce information loss; and a support vector in an SVR is used to find better interfaces, so that classification planes are better, and the generalization capability is higher.

Description

A kind of apparent age method of estimation based on convolutional neural networks and device

Technical field

The invention belongs to image identification technical field is and in particular to a kind of apparent age based on convolutional neural networks is estimated Method and device.

Background technology

Although the automatic age is estimated particularly significant, compared with the phase research work of the face such as Expression Recognition, recognition of face, certainly Dynamic age estimation technique is started late.Initial Research Literature can trace back to the paper that Kwon and Lobo in 1994 delivers, very Research in positive meaning can trace back to the work of Lanities in 2002.So far, no matter at home or external, concern The research or relatively small number of that age is estimated.

Age estimation is regression problem, and conventional linear returns and SVR is linear model, can only process the feelings of linear separability Condition.Even if the SVR plus kernel function can process the situation of linearly inseparable, but kernel function is the manual fixation setting, Can not be poor from mass data learning, generalization ability.

In view of drawbacks described above, creator of the present invention passes through long research and practice obtains the present invention finally.

Content of the invention

For solving above-mentioned technological deficiency, the technical solution used in the present invention is, provides one kind to be based on convolutional neural networks Apparent age method of estimation, it includes：

Step a, Face datection, determine face picture size；

Step b, Image semantic classification, the RGB average of each pixel is calculated on training set；

Step c, the RGB average of each pixel is input to many convolutional layers；

Step d, enters pond layer, carries out pondization operation；

Step e, enters full articulamentum, adjacent two layers is connected entirely by weights；

Step f, supervises structure using softmax loss function depth；

Step g, reduces information using time delay reverse conduction and loses；

Step h：Carry out returning supervision using hinge loss function, the information that reduces is lost；

Step i：Probability matrix.

Preferably, in described step f, the expression formula of softmax function is：

α_i=∑_kh_kW_ki

In above formula, h represents the excitation function of node layer second from the bottom, and W represents and connects layer second from the bottom and softmax layer Weights.

Preferably, in described step g, information is reduced using the number of plies reducing conduction losing.

Preferably, hinge loss regression function is in described step h：

In feature space F, W is one of F vector, and input x is mapped as one of F vector, different in order to process Constant value introduces slack variable ξ_i,Optimize following formula and solve W and b：

${\min }_{\mathrm{\ω},b}P=\frac{1}{2}{W}^{T}W+{\mathrm{C\Σ}}_{i=1}^l({\mathrm{\ξ}}_i+{\mathrm{\ξ}}_i^{*})$

5th, a kind of apparent age estimation unit based on convolutional neural networks, it includes：

Face detection module adopts human-face detector detection original image and discrete postrotational image, correctly demarcates face, Face picture size normalization is 256 × 256 pixels, as the input picture of follow-up Tracking Recognition；

Image pre-processing module, calculates the RGB average of each pixel on training set；

Pond layer module 4, is reduced the characteristic vector of convolutional layer output, improves result simultaneously, be less prone to by pond Over-fitting：

Full articulamentum module 5, adjacent two layers are connected entirely by weights；

Softmax loss module, carries out depth supervision structure using softmax function；

Module is conducted in time delay, using time delay conduction, reduces the number of plies of conduction, reduces confidence loss；

Hinge loss module, using returning loss function hinge loss, reverse conduction to preceding layer, minimizing information is lost Lose；

Probability matrix module, the output of full articulamentum makes softmax layer obtain the probability matrix that this sample belongs to every class, Reach the classification of this test image by the most probable value in probability matrix.

Preferably, described softmax loss module using the expression formula of softmax function is：

α_i=Σ_kh_kW_ki

In above formula, h represents the excitation function of node layer second from the bottom, and W represents and connects layer second from the bottom and softmax layer Weights.

Preferably, described time delay reverse conduction module reduces information using the number of plies reducing conduction losing.

Preferably, described hinge loss module using hinge loss regression function is：

In feature space F, W is one of F vector, and input x is mapped as one of F vector, different in order to process Constant value introduces slack variable ξ_i,Optimize following formula and solve W and b：

${\min }_{\mathrm{\ω},b}P=\frac{1}{2}{W}^{T}W+{\mathrm{C\Σ}}_{i=1}^l({\mathrm{\ξ}}_i+{\mathrm{\ξ}}_i^{*})$

Compared with the prior art the beneficial effects of the present invention is：By depth prison is carried out using softmax loss function Superintend and direct structure；Time delay transmission mechanism reduces conducting shell to reduce information loss；Hinge loss module reduces conducting shell, decreases letter Breath is lost；More preferable interface can be found using the supporting vector in SVR so that classification plane is more excellent, generalization ability is higher.

Brief description

For the technical scheme being illustrated more clearly that in various embodiments of the present invention, below will be to required in embodiment description The accompanying drawing using is briefly described.

Fig. 1 is a kind of flow chart of apparent age method of estimation based on convolutional neural networks of the embodiment of the present invention one；

Fig. 2 is a kind of function of apparent age estimation unit being based on convolutional neural networks of the embodiment of the present invention 6 Schematic diagram.

Specific embodiment

Below in conjunction with accompanying drawing, the above-mentioned He other technical characteristic of the present invention and advantage are described in more detail.

Embodiment one

Fig. 1 is a kind of flow chart of the apparent age method of estimation based on convolutional neural networks of the present invention, wherein, described base Apparent age method of estimation in convolutional neural networks includes：

Step a：Face datection, determines face picture size；

Step b：Image semantic classification, calculates the RGB average of each pixel on training set；

Step c：The RGB average of each pixel is input to many convolutional layers；

Step d：Pond layer, enters pond layer, carries out pondization operation；

Step e, enters full articulamentum, adjacent two layers is connected entirely by weights；

Step f：Supervise structure using softmax loss function depth；

Step g：Reduce information using time delay reverse conduction to lose；

Step h：Carry out returning supervision using hinge loss function, the information that reduces is lost；

Step i：Probability matrix.

The having the beneficial effects that of apparent age method of estimation based on convolutional neural networks：Knot in SVR and linear regression Structure front end adds CNN depth network structure, from mass data learning to rule.Because the supporting vector in SVR can find Preferably, so that classification plane is more excellent, generalization ability is higher for interface.

Embodiment two

Apparent age method of estimation based on convolutional neural networks as above, what the present embodiment was different from is in In, in step a, Face datection is the most basic part of whole face system.For all of training and test pictures, adopt Obtain the accurate location of face with human-face detector.For more accurate demarcation face, in original image and postrotational image all Human-face detector is taken to obtain more preferable effect.Limited due to computer resource, using the image of discrete rotation, polygonal In the degree rotation testing result that obtains of image, select the facial image of highest scoring, and rotated to face to. If can not find human face target in facial image, just taking entire image to be trained and testing.After detecting face, upper Lower left and right four direction extension facial size, increase information is to obtain more preferable effect.If face is included enough Information, directly carries out 0 filling in boundary, guarantees that the face picture size detecting is identical with artwork piece with this.Last people Face picture size is normalized to as 256 × 256 pixels, as the input picture of follow-up Tracking Recognition.

Embodiment three

Apparent age method of estimation based on convolutional neural networks as above, what the present embodiment was different from is in In, in step f after Quan Lian stratum using softmax loss function carry out depth supervision structure, softmaxloss function is such as Following formula, in formula, h represents the excitation function of node layer second from the bottom, and W represents the weights connecting layer second from the bottom and softmax layer, α Represent softmax function：

α_i=Σ_kh_kW_ki

Then,

$p_i=\frac{\exp \left({\mathrm{\α}}_i\right)}{{\mathrm{\Σ}}_j^l\exp \left({\mathrm{\α}}_i\right)}$

Prediction classification be：

I=arg max_ip_i=arg max_iα_i

In order to improve Forecasting recognition rate, calculate the desired value of softmax probability,

$E\left(O\right)={\mathrm{\Σ}}_{i=0}^n{y}_i{o}_i$

Wherein, O=0,1 ..., n represent the age dimension of n+1 output layer, o_iRepresent the output probability of softmax, y_iTable Show the discrete age value of i-th classification, E represents the discrete age value of this sample predictions.

Example IV

Apparent age method of estimation based on convolutional neural networks as above, what the present embodiment was different from is in In in step g, time delay reverse conduction mechanism principle is：Although deeper network has higher levels of ability to express, deeper Neutral net may not necessarily obtain more preferable expression effect, or even decreases, and only pass through increase network depth and may damage Whole estimated performance；Theoretical according to Information Conduction, Information Conduction process can lead to information loss, and the number of plies of reverse conduction is more, More information losses can be caused.Lose to reduce information, introduce the number of plies that time delay conduction reduces conduction, limit reverse conduction In error only propagate a number of number of plies；Time delay is conducted through and introduces extra supervisory signals to carry out reverse conduction, often Individual supervisory signals are only responsible for propagating backward to certain number of plies just stopping, then by another one supervisory signals relay, anti-to bottom To propagated error, continue according to this；Overlap between two adjacent supervisory signals, have the layer of overlap, just using weighting Average mode calculates gradient, does not have the layer of overlap, just directly adopts calculated conduction gradient, so, by boarding steps Degree decline just can solve to whole network.

Embodiment five

Apparent age method of estimation based on convolutional neural networks as above, the present embodiment is different from part and is Using loss regression function Hinge Loss supervision time delay conducting path in step h.

SVR refers to meet the regression function of successive value, and its loss function is Hinge Loss.Hinge Loss regression function For：

In feature space F, W is one of F vector, and input x is mapped as one of F vector.Different in order to process Constant value introduces slack variable ζ_i,Optimize following equations and solve W and b：

In feature space F, W is one of F vector, and input x is mapped as one of F vector, different in order to process Constant value introduces slack variable ξ_i,Optimize following formula and solve W and b：

${\min }_{\mathrm{\ω},b}P=\frac{1}{2}{W}^{T}W+{\mathrm{C\Σ}}_{i=1}^l({\mathrm{\ξ}}_i+{\mathrm{\ξ}}_i^{*})$

Input pattern is mapped to feature space by mapping φ by SVR structure, then by the training mode after mapping and core Carry out point multiplication operation, then different items is weighted suing for peace.Finally increase constant term b come to obtain final predict the outcome defeated Go out.Information in order to avoid time delay reverse conduction is lost, and a recurrence supervision reverse conduction, to preceding layer, returns supervision using support Vector regression is supervised.

Embodiment six

The present embodiment is a kind of apparent age estimation unit based on convolutional neural networks, and it is with described based on convolution god Apparent age method of estimation through network is corresponding, as shown in Fig. 2 it includes：

Face detection module, detects original image and discrete postrotational image using human-face detector, correctly demarcates face, Face picture size normalization is 256 × 256 pixels, as the input picture of follow-up Tracking Recognition；

Image pre-processing module, calculates the RGB average of each pixel on training set；

Many convolutional layers module, the RGB average of each pixel is input to many convolutional layers：

Pond layer module 4, is reduced the characteristic vector of convolutional layer output, improves result simultaneously, be less prone to by pond Over-fitting：

Full articulamentum module 5, adjacent two layers are connected entirely by weights；

Softmax loss module 6, carries out depth supervision structure using softmax function；

Module 7 is conducted in time delay, using time delay conduction, reduces the number of plies of conduction, reduces confidence loss；

Hinge loss module 8, using returning loss function hinge loss, reverse conduction to preceding layer, reduces information Lose；

Probability matrix module 9, the output of full articulamentum makes softmax layer obtain the probability square that this sample belongs to every class Battle array, reaches the classification of this test image by the most probable value in probability matrix.

Supervise structure by depth is carried out using softmax loss function；Time delay transmission mechanism reduces conducting shell to reduce Information loss；Hinge loss module reduces conducting shell, and the information that decreases is lost；Can be found using the supporting vector in SVR Preferably, so that classification plane is more excellent, generalization ability is higher for interface.

The foregoing is only presently preferred embodiments of the present invention, be merely illustrative for the purpose of the present invention, and non-limiting 's.Those skilled in the art understands, it can be carried out in the spirit and scope that the claims in the present invention are limited with many changes, Modification, in addition equivalent, but fall within protection scope of the present invention.

Claims (8)

1. a kind of apparent age method of estimation based on convolutional neural networks is it is characterised in that include：

Step a, Face datection, determine face picture size；

Step b, Image semantic classification, the RGB average of each pixel is calculated on training set；

Step c, the RGB average of each pixel is input to many convolutional layers；

Step d, enters pond layer, carries out pondization operation；

Step e, enters full articulamentum, adjacent two layers is connected entirely by weights；

Step f, supervises structure using softmax loss function depth；

Step g, reduces information using time delay reverse conduction and loses；

Step h：Carry out returning supervision using hinge loss function, the information that reduces is lost；

Step i：Probability matrix.

2. the apparent age method of estimation based on convolutional neural networks according to claim 1 is it is characterised in that described step In rapid f, the expression formula of softmax function is：

α_i=∑_kh_kW_ki

In above formula, h represents the excitation function of node layer second from the bottom, and W represents the weights connecting layer second from the bottom and softmax layer.

3. the apparent age method of estimation based on convolutional neural networks according to claim 2 is it is characterised in that described step Reduce information using the number of plies reducing conduction in rapid g to lose.

4. the face characteristic recognition methods based on SVMs according to claim 3 is it is characterised in that described step In h, hinge loss regression function is：

In feature space F, W is one of F vector, input x is mapped as one of F vector, in order to process exceptional value Introduce slack variable ξ_i,Optimize following formula and solve W and b：

${\min }_{\mathrm{\ω},b}P=\frac{1}{2}{W}^{T}W+{\mathrm{C\Σ}}_{i=1}^l({\mathrm{\ξ}}_i+{\mathrm{\ξ}}_i^{*})$

5. a kind of described apparent age method of estimation corresponding dress based on convolutional neural networks arbitrary with claim 1-4 Put it is characterised in that described included based on the face characteristic identifying device of SVMs：

Face detection module adopts human-face detector detection original image and discrete postrotational image, correctly demarcates face, face Picture size is normalized to 256 × 256 pixels, as the input picture of follow-up Tracking Recognition；

Image pre-processing module, calculates the RGB average of each pixel on training set；

Many convolutional layers module, the RGB average of each pixel is input to many convolutional layers：

Pond layer module 4, is reduced the characteristic vector of convolutional layer output, improves result simultaneously, be less prone to plan by pond Close：

Full articulamentum module 5, adjacent two layers are connected entirely by weights；

Softmax loss module, carries out depth supervision structure using softmax function；

Module is conducted in time delay, using time delay conduction, reduces the number of plies of conduction, reduces confidence loss；

Hinge loss module, using returning loss function hinge loss, reverse conduction to preceding layer, the information that reduces is lost；

Probability matrix module, the output of full articulamentum makes softmax layer obtain the probability matrix that this sample belongs to every class, passes through Most probable value in probability matrix reaches the classification of this test image.

6. the apparent age estimation unit based on convolutional neural networks according to claim 5 is it is characterised in that described Softmax loss module using the expression formula of softmax function is：

α_i=∑_kh_kW_ki

In above formula, h represents the excitation function of node layer second from the bottom, and W represents the weights connecting layer second from the bottom and softmax layer.

7. the apparent age estimation unit based on convolutional neural networks according to claim 6 is it is characterised in that described prolong When reverse conduction module using reduce conduction the number of plies reduce information lose.

8. the apparent age estimation unit based on convolutional neural networks according to claim 7 is it is characterised in that described Hinge loss module using hinge loss regression function is：

In feature space F, W is one of F vector, input x is mapped as one of F vector, in order to process exceptional value Introduce slack variable ξ_i,Optimize following formula and solve W and b：

${\min }_{\mathrm{\ω},b}P=\frac{1}{2}{W}^{T}W+{\mathrm{C\Σ}}_{i=1}^l({\mathrm{\ξ}}_i+{\mathrm{\ξ}}_i^{*})$