CN112906668B

CN112906668B - Face information identification method based on convolutional neural network

Info

Publication number: CN112906668B
Application number: CN202110375280.4A
Authority: CN
Inventors: 任金鹏; 刘云翔; 原鑫鑫; 熊婷婷; 肖岩
Original assignee: Shanghai Institute of Technology
Current assignee: Shanghai Institute of Technology
Priority date: 2021-04-07
Filing date: 2021-04-07
Publication date: 2023-08-25
Anticipated expiration: 2041-04-07
Also published as: CN112906668A

Abstract

The invention relates to the technical field of computer vision and discloses a face information identification method based on a convolutional neural network. The invention overcomes the defect that the traditional method cannot solve the influence of position, rotation and the like on the classification effect, and can more accurately predict the gender and age of the person through the face.

Description

Face information identification method based on convolutional neural network

Technical Field

The invention relates to the technical field of computer vision, in particular to a face information identification method based on a convolutional neural network.

Background

At present, leading edge technologies such as deep learning, artificial intelligence and the like are continuously integrated into the daily life of people, and even vending machines which are small enough to be seen everywhere in life are started to use modes such as face brushing payment and the like. The detection and recognition of faces are not far from reach, but are still in a relatively blank stage in the research field according to the gender and age of face recognition. Faces possess a large amount of useful information, and these information are stable, natural and unique, which are effective ways to identify a person. Meanwhile, the face recognition technology is also an indispensable component of human-computer information interaction, and because the technology does not need to be contacted, the degree of user friendliness is greatly improved, so that the face recognition technology in recent years has a great specific gravity in the market and is rapid in development.

Based on the information identification of the human face, the extraction of the information characteristics of the human face is the precondition of the human face identification by acquiring the picture or video of the human face, the accuracy of the information characteristic extraction can greatly influence the effect of the human face identification, and conversely, the human face identification method has a wide application scene for carrying out the human face identification on the picture of the human face on the basis of high accuracy of the human face identification.

Disclosure of Invention

The invention provides a face information identification method based on a convolutional neural network aiming at the blank existing in face gender and age judgment in face identification.

In order to solve the technical problems, the technical scheme of the invention is as follows:

a face information identification method based on convolutional neural network comprises the following steps:

s1, collecting face pictures, classifying and marking the face pictures, storing the marked results in corresponding CSV files, and sorting and storing the face pictures according to age and gender classification to obtain an Adience face data set;

s2, preprocessing a face picture in the face data set of the science, judging whether the picture has a face or not according to a pre-trained face detection model MTCNN, and marking the face position if the picture has the face;

s3, cutting off the detected face area, and then enhancing the blurred image by utilizing color histogram equalization;

s4, selecting pictures with labels meeting a specified value interval according to the requirement for each face picture, for example, discarding the non-conforming face pictures in a specified age range;

s5, carrying out a series of random operations including cutting, rotating and overturning on the screened face pictures, so that limited picture data can generate more equivalent picture data;

s6, designing a face recognition convolutional neural network model for the face picture, the gender and the age label, and naming the face recognition convolutional neural network model as Remon_Net;

s7, inputting the processed face picture data and the calibrated face information classification result into a designed face recognition convolutional neural network model Remon_Net for parameter training until convergence;

s8, carrying out face information prediction on the video input stream by using a trained face recognition convolutional neural network model Remon_Net, and carrying out frame selection on targets belonging to a predetermined category to obtain a face region and category information.

Preferably, the step S2 includes the following steps:

s21, using the existing Adience face data set, integrating the data, removing invalid files and integrating tag files;

s22, building an existing face detection model MTCNN, and inputting the integrated face data set into the model for training until parameters converge to obtain the trained face detection model MTCNN;

s23, applying the trained face detection model to the Adience face data set in the step S1.

Preferably, the step S3 includes the following steps:

s31, intercepting an original face picture including a hair part for the detected face region;

s32, enhancing the partially blurred image by adopting a color image histogram equalization method;

s33, for the histogram equalization of the color image in the step S32, when the nonlinear extension of the image occurs, the pixel values are redistributed, and the finally output image is a discontinuous histogram with smooth top end;

and S34, carrying out equalization processing on three channels of the color image, and finally merging and outputting.

Preferably, the step S6 includes the steps of:

s61, designing eight layers of convolutions of a face recognition convolutional neural network model Remon_Net, extracting basic features from the first four layers of convolutions, and extracting advanced features from the last four layers of convolutions;

s62, selecting a convolution kernel with a small size for the eight-layer convolution;

s63, adopting a Relu function for the activation function between the convolutions of each layer, and using a learning rate adjustment function to monitor the change of the loss value so as to adjust the learning rate;

s64, the output of the characteristic extraction network is shared to two classifiers, namely an age classifier and a gender classifier, the age classifier finally outputs an age interval, and the gender classifier outputs the gender label of the face picture.

Preferably, the step S8 includes the steps of:

s81, receiving unknown face pictures or video stream data acquired by a camera, intercepting the video stream data according to time frames, and storing data records of a time sequence into a computer for input;

s82, for the image data, using a trained face detection model MTCNN to detect the face and regress the face area, discarding the image data if the face is not detected, otherwise, storing the data;

s83, inputting all face areas in the picture into a trained face recognition convolutional neural network model Remon_Net for prediction, and outputting face information data of each face area;

and S84, integrating all detection results, corresponding to the position of each face region, including the attributes of the face size, the age and the sex, and returning and outputting the real time in a video stream window.

Compared with the prior art, the invention has the beneficial effects that:

according to the invention, the original face image is firstly subjected to processing such as segmentation and alignment, the influence of the background and the angle on the classification effect is avoided, the segmented and aligned face data is input through a designed network model, an eight-layer convolutional neural network is used for the feature extraction part, and then two fully connected modules are used as classifiers for gender identification and age identification. The invention adopts the convolutional neural network model to extract the characteristics, and overcomes the defect that the traditional method can not solve the influence of position, rotation and the like on the classification effect, so the method provided by the invention can more accurately predict the gender and age of the person through the face. Meanwhile, the output of the face feature extraction network is directly shared to two parallel classifier modules in the face recognition convolutional neural network model remote_Net, so that redundant calculation amount and overlong prediction time caused by classification work by using two independent models are avoided, and the problem of low real-time performance of the traditional image classification in face information recognition is solved to a certain extent.

Of course, it is not necessary for any one product to practice the invention to achieve all of the advantages set forth above at the same time.

Drawings

FIG. 1 is a flow chart of a face information recognition method based on convolutional neural network;

fig. 2 is a schematic diagram of a face recognition convolutional neural network model remon_net of the present invention.

Detailed Description

The following describes the embodiments of the present invention further with reference to the drawings. The description of these embodiments is provided to assist understanding of the present invention, but is not intended to limit the present invention. In addition, the technical features of the embodiments of the present invention described below may be combined with each other as long as they do not collide with each other.

Referring to fig. 1, a face information recognition method based on a convolutional neural network includes the steps of:

The accuracy of the trained face detection model can reach more than 94%.

the face region is cut, so that complex backgrounds can be removed, effective characteristics of pictures are reserved, and redundant characteristics are reduced.

in order to keep more picture information, the detection result is expanded to a certain extent, so that the hair information can be kept.

the main body features are made more obvious and clear by enhancing the partially blurred image in turn.

under certain conditions, the number of pixel gray levels which are different from each other tends to be balanced, and the contrast of the part which is not processed is more obvious, so that the finally output image is a discontinuous histogram with smooth top.

more equivalent picture data can prevent the over-fitting problem in the model training process, namely, the method is effective only for selecting the trained face picture and is ineffective for unknown face recognition.

the basic features are local features such as an organ, and the advanced features are texture lines of the organ.

the eight-layer convolution can realize the functions of feature extraction and classification, and small-size convolution kernels, such as 3×3, are selected, so that the calculated amount in the training process can be reduced, and the face recognition rate is accelerated.

the Relu function is used mainly to add non-linearity to the model, which is not well expressed. The Relu function judges whether the output of the current node reaches a threshold value, and the output is transferred to the next layer if the output reaches the threshold value, and the output is not transferred if the output does not reach the threshold value.

Referring to fig. 2, the face recognition convolutional neural network model remon_net includes a left feature extraction network and two classification modules on the right. The feature extraction network includes eight layers, each having a convolutional layer, a regularization layer, and an activation function layer, and the underlying data represents the number of channels of the convolutional layer. The two classifier modules are connected by the full connection layer because the output is flattened to one dimension as well as downsampling by a maximum pooling layer through the output of the feature extraction network, and the following numbers represent the node number of the full connection layer. The last is the output classification book, gender is 2-class, and age is 8-class, i.e., the likelihood of outputting gender and the likelihood of each age group.

"framing targets belonging to a predetermined category" means that only faces are detected, and information categories of the faces are discriminated so as to be in line with the faces and the categories are framed in the detection effect only when they are within the range.

S81, receiving video stream data acquired by a camera, intercepting the video stream data according to time frames, and storing a data record of a time sequence into a computer memory for input, wherein the data record corresponds to each input data, namely image data of each frame;

The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. It will be apparent to those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, and yet fall within the scope of the invention.

Claims

1. The face information identification method based on the convolutional neural network is characterized by comprising the following steps of:

s4, selecting pictures with labels meeting a specified value interval according to the requirement for each face picture, and discarding non-conforming face pictures;

s8, carrying out face information prediction on the video input stream by using a trained face recognition convolutional neural network model Remon_Net, and carrying out frame selection on targets belonging to a predetermined category to obtain a face region and category information;

the step S2 includes the following steps:

s23, applying the trained face detection model to the Adience face data set in the step S1;

step S3 comprises the steps of:

s34, carrying out equalization treatment on three channels of the color image, and finally merging and outputting;

step S6 includes the steps of:

s64, sharing the output of the characteristic extraction network to two classifiers, namely an age classifier and a gender classifier, wherein the age classifier finally outputs an age interval, and the gender classifier outputs the gender label of the face picture;

step S8 includes the steps of: