WO2021012383A1

WO2021012383A1 - Age prediction method and device for infrared image

Info

Publication number: WO2021012383A1
Application number: PCT/CN2019/108078
Authority: WO
Inventors: 吴梓恒; 胡杰
Original assignee: 初速度（苏州）科技有限公司
Priority date: 2019-07-24
Filing date: 2019-09-26
Publication date: 2021-01-28
Also published as: CN112307796B; CN112307796A

Abstract

An age prediction method and device for an infrared image. The method comprises: acquiring a to-be-processed infrared image (S110); detecting a first human face area in the to-be-processed infrared image, and constructing a to-be-processed target image containing the first human face, the size of the target image being a preset size (S120); inputting the to-be-processed target image into an infrared convolutional neural network model obtained by pre-training so as to obtain a first predicted age distribution of a person corresponding to the first human face area, the first predicted age distribution following the Gaussian distribution. The infrared convolutional neural network model is obtained by adjusting all parameters in an initial infrared convolutional neural network model after all infrared sample images are input into the initial infrared convolutional neural network model and according to a difference between an age distribution corresponding to all infrared sample images output by the initial infrared convolutional neural network model and a Gaussian distribution generated by a corresponding age annotation result and a difference between the desired value of the age distribution corresponding to all infrared sample images and the corresponding age annotation result, the age distribution following the Gaussian distribution. The infrared sample images and the corresponding age annotation result are determined on the basis of the convolutional neural network model obtained by pre-training of color images (S130). The described solution can be used for performing age prediction to all infrared images.

Description

An age prediction method and device for infrared images

Technical field

The present invention relates to the technical field of image processing, in particular to an age prediction method and device for infrared images.

Background technique

At present, age prediction based on surveillance images mainly uses convolutional neural networks. Specifically, firstly, a convolutional neural network model needs to be trained through sample images and accurate age annotation results, and then the age prediction of the face in the predicted image can be performed based on the trained convolutional neural network model.

However, the existing age-labeled image set is a color image set, and the convolutional neural network model trained based on the image set can only predict the age of the face in the color image. For infrared images, due to the lack of an age-labeled data set, it is impossible to train a convolutional neural network model that predicts the age of the face in the infrared image. Therefore, in order to predict the age of the face in the infrared image, an age prediction method for the infrared image is urgently needed.

Summary of the invention

The present invention provides an age prediction method and device for infrared images to predict the age of human faces in infrared images. The specific technical solution is as follows.

In the first aspect, an embodiment of the present invention provides an age prediction method for infrared images, and the method includes:

Obtain infrared images to be processed;

Detecting a first face region in the infrared image to be processed, and constructing a target image to be processed including the first face region; wherein the size of the target image to be processed is a preset size;

Inputting the target image to be processed into a pre-trained infrared convolutional neural network model to obtain a first predicted age distribution of a person corresponding to the first face region, wherein the first predicted age distribution obeys a Gaussian distribution;

Wherein, the infrared convolutional neural network model is based on the input of each infrared sample image to the initial infrared convolutional neural network model, the initial infrared convolutional neural network model outputs the corresponding age distribution of each infrared sample image and the corresponding age annotation result The difference between the generated Gaussian distribution and the difference between the expected value of the age distribution corresponding to each infrared sample image and the corresponding age annotation result are obtained after adjusting each parameter in the initial infrared convolutional neural network model, the age The distribution obeys the Gaussian distribution; the infrared sample image and the corresponding age annotation result are determined according to a pre-trained convolutional neural network model, and the convolutional neural network model is obtained through color image training.

Optionally, the training process of the infrared convolutional neural network model includes:

Construct an initial infrared convolutional neural network model, the initial infrared convolutional neural network model includes: a convolutional layer, a pooling layer, and a fully connected layer;

Determine each infrared sample image and the age annotation result corresponding to each infrared sample image;

Generating a Gaussian distribution of the age annotation result corresponding to each infrared sample image;

Input each infrared sample image into the initial infrared convolutional neural network model to obtain the age distribution corresponding to each infrared sample image, and calculate the age distribution corresponding to each infrared sample image and the Gaussian distribution generated by the corresponding age annotation result The difference between the expected value of the age distribution corresponding to each infrared sample image and the difference between the corresponding age annotation results, and the parameters in the initial infrared convolutional neural network model are adjusted according to the calculation results to obtain the infrared convolutional neural network Network model.

Optionally, the determining each infrared sample image and the age marking result corresponding to each infrared sample image includes:

Acquiring a plurality of image sets, wherein the initial infrared images in each image set are different facial images of the same person in the same period, and the number of initial infrared images in each image set is greater than a preset number threshold;

For each of the image sets, detecting the second face region in each initial infrared image, and constructing each initial target image including each of the second face regions;

Input each of the initial target images into a pre-trained convolutional neural network model to obtain the second predicted age distribution of the person corresponding to each of the second face regions, and determine the age range corresponding to each second predicted age distribution; Wherein, the convolutional neural network model is based on the input of the initial convolutional neural network model of each sample image, the age distribution corresponding to each sample image output by the initial convolutional neural network model and the Gaussian distribution generated by the corresponding age annotation result The difference, and the difference between the expected value of the age distribution corresponding to each sample image and the corresponding age annotation result, adjust each parameter in the initial convolutional neural network model to obtain a candidate neural network model, and compare the candidate neural network model Obtained after adjustment, the second predicted age distribution obeys a Gaussian distribution, and each sample image is a color image;

For each image set, remove the initial target images with abnormal age ranges in the image set to obtain the remaining target images, calculate the normal age range corresponding to all the remaining target images, and use the remaining target images included in the normal age range as Infrared sample images, and the average value of the age range corresponding to each infrared sample image is used as the age labeling result of each infrared sample image.

Optionally, for each image set, removing initial target images with abnormal age ranges in the image set to obtain the remaining target images includes:

For each image set, sort the initial target images according to the minimum value of the age range corresponding to each initial target image included in the image set in descending order;

Determining a first age range located at one quarter and a second age range located at three quarters, as well as the minimum value of the first age range and the maximum value of the second age range;

Remove the initial target image whose age value is less than the difference between the minimum value and the preset value in the corresponding age range and the initial target image whose age value is greater than the sum of the maximum value and the preset value to obtain the remaining target image .

Optionally, the calculation of the normal age range corresponding to all remaining target images includes:

Calculate the mean and standard deviation of the age range corresponding to all remaining target images;

Get preset hyperparameters;

Calculate the product of the hyperparameter and the standard deviation, and use the difference between the mean and the product as the minimum value in the normal age range, and use the sum of the mean and the product as the maximum in the normal age range value.

Optionally, the training process of the convolutional neural network model includes:

Construct an initial convolutional neural network model, the initial convolutional neural network model including: a convolutional layer, a pooling layer, and a fully connected layer;

Acquiring each sample image and the age annotation result corresponding to each sample image;

Generating a Gaussian distribution of the age annotation result corresponding to each sample image;

Input each sample image into the initial convolutional neural network model to obtain the age distribution corresponding to each sample image, and calculate the difference between the age distribution corresponding to each sample image and the Gaussian distribution generated by the corresponding age annotation result, As well as the difference between the expected value of the age distribution corresponding to each sample image and the corresponding age annotation result, the parameters in the initial convolutional neural network model are adjusted according to the calculation result to obtain the candidate neural network model, and the candidate neural network The model is adjusted to obtain the convolutional neural network model.

Optionally, the generating the Gaussian distribution of the age annotation result corresponding to each sample image includes:

For each sample image, construct a Gaussian distribution centered on the age annotation result corresponding to the sample image, and the preset standard deviation is the peak width, as the Gaussian distribution of the age annotation result corresponding to the sample image.

Optionally, after the infrared convolutional neural network model is obtained, the method further includes:

Acquiring an infrared test image and an age marking result corresponding to each of the infrared test images; the infrared test image is different from the infrared sample image;

Determining the test accuracy of the infrared convolutional neural network model according to the infrared test image and the age annotation result corresponding to each of the infrared test images;

When the test accuracy is less than the preset accuracy threshold, use the current infrared convolutional neural network model as the initial infrared convolutional neural network model, and return to execute the determination of each infrared sample image and the age corresponding to each infrared sample image In the step of marking the results, until the test accuracy is not less than the preset accuracy threshold, the current infrared convolutional neural network model is used as the final infrared convolutional neural network model.

Optionally, the constructing the to-be-processed target image including the first face region includes:

Performing key point detection on the first face area to obtain coordinate information of each target key point in the first face area; wherein each target key point is a point that identifies a face contour feature;

According to the coordinate information of each target key point, after the first face region is aligned, a target image to be processed including the first face region and each target key point is located at a preset position is obtained.

Optionally, after the input of the target image to be processed into the pre-trained infrared convolutional neural network model, and the first predicted age distribution of the person corresponding to the first face region is obtained, the method further includes:

Calculate the sum of the product of each age value and the corresponding probability in the first predicted age distribution, and use the calculation result as the predicted age value of the person corresponding to the first face area.

In a second aspect, an embodiment of the present invention provides an age prediction device for infrared images, the device includes:

Infrared image acquisition module for acquiring infrared images to be processed;

The face area detection module is used to detect the first face area in the infrared image to be processed, and construct a target image to be processed containing the first face area; wherein the size of the target image to be processed is Preset size

The age prediction module is used to input the target image to be processed into a pre-trained infrared convolutional neural network model to obtain the first predicted age distribution of the person corresponding to the first face region, wherein the first prediction The age distribution obeys the Gaussian distribution; wherein, the infrared convolutional neural network model is the age corresponding to each infrared sample image output by the initial infrared convolutional neural network model after inputting the initial infrared convolutional neural network model according to each infrared sample image The difference between the distribution and the Gaussian distribution generated by the corresponding age annotation result, and the difference between the expected value of the age distribution corresponding to each infrared sample image and the corresponding age annotation result, after adjusting the parameters in the initial infrared convolutional neural network model It is obtained that the age distribution obeys the Gaussian distribution; the infrared sample image and the corresponding age annotation result are determined according to a pre-trained convolutional neural network model, which is obtained through color image training.

Optionally, the device further includes:

The infrared model building module is used to build an initial infrared convolutional neural network model, the initial infrared convolutional neural network model includes: a convolutional layer, a pooling layer, and a fully connected layer;

An infrared sample image determination module, used to determine each infrared sample image and the age annotation result corresponding to each infrared sample image;

A Gaussian distribution generating module, configured to generate the Gaussian distribution of the age annotation result corresponding to each infrared sample image;

The infrared convolutional neural network model training module is used to input each infrared sample image into the initial infrared convolutional neural network model, obtain the age distribution corresponding to each infrared sample image, and calculate the corresponding infrared sample image The difference between the age distribution and the Gaussian distribution generated by the corresponding age labeling result, and the difference between the expected value of the age distribution corresponding to each infrared sample image and the corresponding age labeling result, according to the calculation result of the initial infrared convolutional neural network model The parameters are adjusted to obtain the infrared convolutional neural network model.

Optionally, the infrared sample image determination module includes:

The image collection acquisition sub-module is used to acquire multiple image collections, wherein the initial infrared images in each image collection are different facial images of the same person in the same period, and the initial infrared images in each image collection The number of is greater than the preset number threshold;

The face area detection sub-module is used to detect the second face area in each initial infrared image for each of the image sets, and construct each initial target image including each of the second face areas;

The age range determination sub-module is used to input each of the initial target images into a pre-trained convolutional neural network model to obtain the second predicted age distribution of the person corresponding to each of the second face regions, and determine each second Predict the age range corresponding to the age distribution; wherein the convolutional neural network model is based on the input of the initial convolutional neural network model of each sample image, and the initial convolutional neural network model outputs the corresponding age distribution of each sample image and the corresponding The difference between the Gaussian distribution generated by the age labeling result, and the difference between the expected value of the age distribution corresponding to each sample image and the corresponding age labeling result, adjusting each parameter in the initial convolutional neural network model to obtain a candidate neural network model, And obtained after adjusting the candidate neural network model, the second predicted age distribution obeys a Gaussian distribution, and each sample image is a color image;

The infrared sample determination sub-module is used to remove the initial target images with abnormal age ranges in the image set for each image set to obtain the remaining target images, calculate the normal age range corresponding to all remaining target images, and include them in the normal The remaining target images within the age range are used as infrared sample images, and the average value of the age range corresponding to each infrared sample image is used as the age labeling result of each infrared sample image.

Optionally, the infrared sample determination sub-module is specifically used for:

Get preset hyperparameters;

Optionally, the infrared sample image determination module further includes:

The network model construction sub-module is used to construct an initial convolutional neural network model, and the initial convolutional neural network model includes: a convolutional layer, a pooling layer, and a fully connected layer;

The sample image acquisition sub-module is used to acquire each sample image and the age annotation result corresponding to each sample image;

The Gaussian distribution generation sub-module is used to generate the Gaussian distribution of the age annotation results corresponding to each sample image;

The convolutional neural network model training sub-module is used to input each sample image into the initial convolutional neural network model, obtain the age distribution corresponding to each sample image, and calculate the age distribution corresponding to each sample image and the corresponding The difference between the Gaussian distribution generated by the age annotation result, and the difference between the expected value of the age distribution corresponding to each sample image and the corresponding age annotation result, according to the calculation result, adjust each parameter in the initial convolutional neural network model to obtain the candidate Neural network model, and adjusting the candidate neural network model to obtain the convolutional neural network model.

Optionally, the Gaussian distribution generating sub-module is specifically used for:

Optionally, the device further includes:

A test image acquisition module, configured to acquire an infrared test image and an age marking result corresponding to each of the infrared test images; the infrared test image is different from the infrared sample image;

A test accuracy determining module, configured to determine the test accuracy of the infrared convolutional neural network model according to the infrared test image and the age annotation result corresponding to each infrared test image;

The processing module is configured to use the current infrared convolutional neural network model as the initial infrared convolutional neural network model when the test accuracy is less than the preset accuracy threshold, and trigger the infrared sample image determination module until the test accuracy is not When it is less than the preset accuracy threshold, the current infrared convolutional neural network model is used as the final infrared convolutional neural network model.

Optionally, the face area detection module includes:

The key point detection sub-module is used to perform key point detection on the first face area to obtain coordinate information of each target key point in the first face area; wherein, each target key point is an identification face Points of contour features;

The target image construction sub-module is used to align the first face region according to the coordinate information of the target key points to obtain the first face region and the target key points are located in the preset Set the position of the target image to be processed.

Optionally, the device further includes:

The age value calculation module is used to calculate the sum of the product of each age value and the corresponding probability in the first predicted age distribution, and use the calculation result as the predicted age value of the person corresponding to the first face area.

It can be seen from the above content that the method and device for predicting the age of a face in an infrared image provided by the embodiments of the present invention can obtain an infrared image to be processed; detect the first face area in the infrared image to be processed, and construct a A target image to be processed in the face region; where the size of the target image to be processed is a preset size; the target image to be processed is input into the pre-trained infrared convolutional neural network model to obtain the first face region corresponding to the person The first predicted age distribution, where the first predicted age distribution obeys the Gaussian distribution; among them, the infrared convolutional neural network model is output from the initial infrared convolutional neural network model based on the input of each infrared sample image The difference between the age distribution corresponding to each infrared sample image and the Gaussian distribution generated by the corresponding age labeling result, and the difference between the expected value of the age distribution corresponding to each infrared sample image and the corresponding age labeling result, compare the initial infrared convolutional neural network model After adjusting the parameters, the age distribution obeys the Gaussian distribution; the infrared sample image and the corresponding age annotation result are determined according to the pre-trained convolutional neural network model, which is trained on the color image, so it can The convolutional neural network model obtained by color image training determines the infrared sample image and the corresponding age annotation result, and then trains the determined infrared sample image and the corresponding age annotation result to obtain the age prediction of the face in the infrared image Infrared convolutional neural network model. In addition, compared with manual age calibration, determining infrared sample images and corresponding age annotation results through a convolutional neural network model can save human resources and improve the efficiency of sample acquisition. In addition, when training the infrared convolutional neural network model, after each infrared sample image is input to the initial infrared convolutional neural network model, the initial infrared convolutional neural network model outputs the corresponding age distribution of each infrared sample image and the corresponding age annotation results are generated The difference between the Gaussian distribution of each infrared sample image and the difference between the expected value of the age distribution corresponding to each infrared sample image and the corresponding age annotation result. The parameters in the initial infrared convolutional neural network model are adjusted, which are the same as the output specific age value. In comparison, it can accurately predict the age of the same person from multiple angles and multiple states, and improve the robustness of the model. Of course, implementing any product or method of the present invention does not necessarily need to achieve all the advantages described above at the same time.

The innovative points of the embodiments of the present invention include:

1. Determine the infrared sample image and the corresponding age annotation result based on the convolutional neural network model obtained by color image training, and then train according to the determined infrared sample image and the corresponding age annotation result to be able to age the face in the infrared image The predicted infrared convolutional neural network model. In addition, compared with manual age calibration, determining infrared sample images and corresponding age annotation results through a convolutional neural network model can save human resources and improve the efficiency of sample acquisition. In addition, when training the infrared convolutional neural network model, after each infrared sample image is input to the initial infrared convolutional neural network model, the initial infrared convolutional neural network model outputs the corresponding age distribution of each infrared sample image and the corresponding age annotation results are generated The difference between the Gaussian distribution of each infrared sample image and the difference between the expected value of the age distribution corresponding to each infrared sample image and the corresponding age annotation result. The parameters in the initial infrared convolutional neural network model are adjusted, which are the same as the output specific age value. In comparison, it can accurately predict the age of the same person from multiple angles and multiple states, and improve the robustness of the model.

2. After inputting the initial infrared convolutional neural network model according to each infrared sample image, the difference between the age distribution corresponding to each infrared sample image output by the initial infrared convolutional neural network model and the Gaussian distribution generated by the corresponding age annotation result, and each infrared The difference between the expected value of the age distribution corresponding to the sample image and the corresponding age annotation result. The parameters of the initial infrared convolutional neural network model are adjusted to obtain the infrared convolutional neural network model. Compared with the scheme that outputs the specific age value, it can The same person's multi-angle and multi-state have accurate age prediction, which improves the robustness of the model.

3. Determine the infrared sample image and the corresponding age annotation result through the convolutional neural network model. Compared with manual age calibration, it can save human resources and improve the efficiency of sample acquisition.

4. After inputting the initial convolutional neural network model according to each sample image, the difference between the age distribution corresponding to each sample image output by the initial convolutional neural network model and the Gaussian distribution generated by the corresponding age annotation result, and the age corresponding to each sample image The difference between the expected value of the distribution and the corresponding age annotation result. The parameters in the initial convolutional neural network model are adjusted to obtain the convolutional neural network model. Compared with the scheme that outputs the specific age value, it can be used for multiple angles and multiple angles of the same person. The state has an accurate age prediction, which improves the robustness of the model.

5. Test the test accuracy of the trained infrared convolutional neural network model through the test image, and when the test accuracy is low, update the infrared convolutional neural network model through the infrared sample image again, so as to ensure the final infrared The test accuracy of the convolutional neural network improves the accuracy of age prediction.

6. Perform key point detection on the face area, and then align the face area to obtain the target image to be processed, which can avoid situations such as side faces in the target image to be processed, thereby ensuring that the face in the target image to be processed is clearer. Improve the accuracy of age prediction.

7. Calculate the specific predicted age value according to the predicted age distribution, so as to obtain the accurate age prediction result.

Description of the drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained from these drawings without creative work.

FIG. 1 is a schematic flowchart of an age prediction method for infrared images according to an embodiment of the present invention;

2 is a schematic diagram of another flow chart of an age prediction method for infrared images according to an embodiment of the present invention;

3 is a schematic diagram of another flow chart of an age prediction method for infrared images according to an embodiment of the present invention;

4 is a schematic diagram of another flow chart of an age prediction method for infrared images according to an embodiment of the present invention;

5 is a schematic diagram of another flow chart of an age prediction method for infrared images according to an embodiment of the present invention;

6 is a schematic diagram of another flow chart of an age prediction method for infrared images according to an embodiment of the present invention;

FIG. 7 is a schematic structural diagram of an age prediction device for infrared images according to an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

It should be noted that the terms "including" and "having" in the embodiments of the present invention and the drawings and any variations thereof are intended to cover non-exclusive inclusions. For example, the process, method, system, product or device that contains a series of steps or units is not limited to the listed steps or units, but optionally also includes unlisted steps or units, or optionally also includes Other steps or units inherent to these processes, methods, products or equipment.

The embodiment of the invention discloses an age prediction method and device for infrared images, which can predict the age of a human face in the infrared image. The embodiments of the present invention will be described in detail below.

FIG. 1 is a schematic flowchart of a method for age prediction of infrared images provided by an embodiment of the present invention. This method is applied to electronic equipment. The method specifically includes the following steps.

S110: Acquire an infrared image to be processed.

The above-mentioned infrared image to be processed is an image containing a human face that needs to be age predicted. For example, the electronic device may receive the infrared image collected by the monitoring device as the infrared image to be processed; or, may receive the infrared image input by the user as the infrared image to be processed, which is not limited in the embodiment of the present invention.

S120: Detect a first face region in the infrared image to be processed, and construct a target image to be processed including the first face region; wherein the size of the target image to be processed is a preset size.

It can be understood that due to reasons such as a large monitoring area of the monitoring device, the infrared image to be processed may include areas other than the face area. When predicting age, other regions may affect the results of age prediction.

Therefore, in the embodiment of the present invention, the electronic device can detect the face area in the infrared image to be processed, which can be referred to as the first face area, and construct the to-be-processed target image containing the first face area. The size of the target image to be processed is a preset size.

For example, the Faster-RCNN (FasterRegion-based Cellular Neural Network) face detection framework can be used to detect the first face region in the infrared image to be processed. Alternatively, any known target detection algorithm may be used to detect the first face region in the infrared image to be processed, which is not limited in the embodiment of the present invention.

S130: Input the target image to be processed into the pre-trained infrared convolutional neural network model to obtain the first predicted age distribution of the person corresponding to the first face area, where the first predicted age distribution obeys the Gaussian distribution; where the infrared volume The product neural network model is the difference between the age distribution corresponding to each infrared sample image output by the initial infrared convolution neural network model and the Gaussian distribution generated by the corresponding age annotation result after inputting the initial infrared convolution neural network model according to each infrared sample image. And the difference between the expected value of the age distribution corresponding to each infrared sample image and the corresponding age annotation result, which is obtained after adjusting the parameters in the initial infrared convolutional neural network model. The age distribution obeys the Gaussian distribution; the infrared sample image and the corresponding age The labeling result is determined based on the pre-trained convolutional neural network model, which is trained on the color image.

In the embodiment of the present invention, an infrared convolutional neural network model for age prediction of a human face in an infrared image can be constructed in advance. Specifically, the convolutional neural network model can be obtained by training based on the color image first, and the convolutional neural network model can perform rough age prediction on the infrared image. Then determine the infrared sample image and the corresponding age annotation result based on the above convolutional neural network model, and then input each infrared sample image into the initial infrared convolutional neural network model, according to the initial infrared convolutional neural network model output corresponding to each infrared sample image The difference between the age distribution and the Gaussian distribution generated by the corresponding age annotation result, and the difference between the expected value of the age distribution corresponding to each infrared sample image and the corresponding age annotation result, after adjusting the parameters in the initial infrared convolutional neural network model Infrared convolutional neural network model.

Gaussian distribution, also known as normal distribution. If the random variable X obeys a normal distribution with a mathematical expectation of μ and a variance of σ^2, it is recorded as N(μ,σ^2). Its expected value μ determines its location, and its standard deviation σ determines the magnitude of the distribution. When μ=0 and σ=1, the normal distribution is the standard normal distribution.

The expected value of the age distribution corresponding to each infrared sample image is the expected value of the Gaussian distribution, that is, the age value with the highest probability in the middle of the age distribution corresponding to each infrared sample image.

After the target image to be processed containing the first face area is obtained, the target image to be processed can be input into the infrared convolutional neural network model, and the infrared convolutional neural network model can output the first face area corresponding to the first person. Forecast age distribution. Among them, the first predicted age distribution obeys Gaussian distribution, that is, obeys normal distribution.

The aforementioned predicted age distribution includes multiple age values and corresponding probability values. Among multiple age values, the probability of the age value in the middle is the largest, and the probability of the age value on both sides decreases sequentially. And, the sum of the probabilities of all age values is 1.

It can be seen from the foregoing that the method for predicting the age of a face in an infrared image provided by the embodiment of the present invention can determine the infrared sample image and the corresponding age annotation result based on the convolutional neural network model obtained by color image training, and then according to The determined infrared sample images and the corresponding age annotation results are trained to obtain an infrared convolutional neural network model that can predict the age of the face in the infrared image. In addition, compared with manual age calibration, determining infrared sample images and corresponding age annotation results through a convolutional neural network model can save human resources and improve the efficiency of sample acquisition. In addition, when training the infrared convolutional neural network model, after each infrared sample image is input to the initial infrared convolutional neural network model, the initial infrared convolutional neural network model outputs the corresponding age distribution of each infrared sample image and the corresponding age annotation results are generated The difference between the Gaussian distribution of each infrared sample image and the difference between the expected value of the age distribution corresponding to each infrared sample image and the corresponding age annotation result. The parameters in the initial infrared convolutional neural network model are adjusted, which are similar to the output specific age value. In comparison, it can accurately predict the age of the same person from multiple angles and multiple states, improving the robustness of the model.

As an implementation of the embodiment of the present invention, as shown in FIG. 2, the training process of the infrared convolutional neural network model of the embodiment of the present invention may include the following steps.

S210: Construct an initial infrared convolutional neural network model. The initial infrared convolutional neural network model includes: a convolutional layer, a pooling layer, and a fully connected layer.

The initial infrared convolutional neural network model in the embodiment of the present invention may include data processing layers with parameters such as a convolution layer, a pooling layer, and a fully connected layer. Among them, the number of convolutional layers, pooling layers, and fully connected layers may be one or more layers, as long as age prediction can be realized, which is not limited in the embodiment of the present invention.

S220: Determine each infrared sample image and the age annotation result corresponding to each infrared sample image.

Determine each infrared sample image and the age annotation result corresponding to each infrared sample image, that is, determine the data set used to train the infrared convolutional neural network model.

In an implementation manner, as shown in FIG. 3, the process of determining each infrared sample image and the age annotation result corresponding to each infrared sample image may include the following steps.

S310: Acquire multiple image sets, where the initial infrared images in each image set are different facial images of the same person in the same period, and the number of initial infrared images in each image set is greater than a preset number threshold.

In the embodiment of the present invention, a large number of infrared face images can be collected, and the infrared sample images that can be used to train the infrared convolutional neural network model are determined.

Specifically, by designing the collection method, for different people, multiple infrared facial images of the same period can be acquired. For example, millions of (such as 5 million, 6 million, 7 million, etc.) infrared face images can be collected, with an average of about 100 per person as the initial infrared image.

The same period mentioned above may be a preset period of time, such as 1 day, 30 days, 60 days, etc., which is not limited in the embodiment of the present invention.

S320: For each image set, detect the second face area in each initial infrared image, and construct each initial target image including each second face area.

For example, the Faster-RCNN face detection framework can be used to detect the second face region in each initial infrared image included in each image set. Alternatively, any known target detection algorithm may be used to detect the second face region in each initial infrared image included in each image set, which is not limited in the embodiment of the present invention.

After detecting the second human face regions in the initial infrared images included in each image set, each initial target image including each second human face region can be constructed.

S330: Input each initial target image into the pre-trained convolutional neural network model to obtain the second predicted age distribution of the person corresponding to each second face area, and determine the age range corresponding to each second predicted age distribution; where, The convolutional neural network model is based on the input of each sample image into the initial convolutional neural network model, the initial convolutional neural network model outputs the corresponding age distribution of each sample image and the difference between the Gaussian distribution generated by the corresponding age annotation result, and each sample The difference between the expected value of the age distribution corresponding to the image and the corresponding age annotation result. The parameters in the initial convolutional neural network model are adjusted to obtain the candidate neural network model, and the candidate neural network model is adjusted. The second predicted age The distribution obeys the Gaussian distribution, and each sample image is a color image.

In the embodiment of the present invention, a convolutional neural network model that can predict the age of a human face in an infrared image can be constructed in advance. Specifically, a color image with an age can be used as a sample image, and after each sample image is input into the initial convolutional neural network model, the age distribution corresponding to each sample image output by the initial convolutional neural network model and the corresponding age annotation result are generated The difference between the Gaussian distribution of each sample image and the difference between the expected value of the age distribution corresponding to each sample image and the corresponding age annotation result. The parameters in the initial convolutional neural network model are adjusted to obtain the candidate neural network model. The candidate neural network model can Predict the age of color images; then adjust the candidate neural network model to obtain a convolutional neural network model that can predict the age of infrared images.

After each initial target image containing each second face area is obtained, each initial target image can be input into the pre-trained convolutional neural network model, and the convolutional neural network model can output the person corresponding to each second face area The second predicted age distribution. Among them, each second predicted age distribution obeys Gaussian distribution, that is, obeys normal distribution.

It can be understood that since infrared images and color images have different characteristics, the accuracy of age predicted by the convolutional neural network model is not particularly high. In the embodiment of the present invention, after obtaining the second predicted age distribution of the person corresponding to each second face region, the age range corresponding to each second predicted age distribution can be determined, that is, the age included in each second predicted age distribution range.

S340: For each image set, remove the initial target images with an abnormal age range in the image set to obtain the remaining target images, calculate the normal age range corresponding to all the remaining target images, and use the remaining target images included in the normal age range as Infrared sample images, and the average value of the age range corresponding to each infrared sample image is used as the age labeling result of each infrared sample image.

It can be understood that for hundreds of images of the same person in the same period, the age prediction results should be the same, that is, for each image set, the age range of each image included therein should be the same. However, in practical applications, due to mixing other people's images into this person's image set, or the influence of different image angles, lighting, etc., the age ranges obtained by these images through step S330 will not be exactly the same.

In the embodiment of the present invention, for each image set, the initial target images with abnormal age ranges in the image set may be removed to obtain the remaining target images.

In an implementation manner, for each image set, the interquartile range method may be used to remove the initial target images with abnormal age ranges in the image set to obtain the remaining target images. Specifically, as shown in FIG. 4, the process may include the following steps.

S410: For each image set, sort the initial target images according to the minimum value of the age range corresponding to each initial target image included in the image set from small to large.

For example, when any image set includes 100 initial target images, and the age range of each initial target image is 15-25 for 10 pictures, 30-40 for 80 pictures, and 40-50 for 10 pictures, you can The initial target images are sorted according to the order of the smallest value of each age range from small to large, that is, the order of 15, 30, and 40.

S420: Determine the first age range located at one quarter and the second age range located at three quarters, as well as the minimum value of the first age range and the maximum value of the second age range.

In the above example, the first age range located at one quarter is the age range 30-40 corresponding to the 25th initial target image, and the second age range located at three quarters is the 75th initial target image The corresponding age range is 30-40. The minimum value of the first age range is 30, and the maximum value of the second age range is 40.

S430: Remove the initial target images whose age values are less than the difference between the minimum value and the preset value in the corresponding age range and the initial target images whose age values are greater than the sum of the maximum value and the preset value to obtain the remaining target images.

The foregoing preset value may be any preset number, such as 3, 5, 6, etc., which is not limited in the embodiment of the present invention.

For example, when the preset value is 3, in the above example, the difference between the minimum value and the preset value is 27, and the sum of the maximum value and the preset value is 43. The age range includes age values less than the minimum value and the preset value. The initial target image with the difference between the values is 10 initial target images with an age range of 15-25. The age range includes the initial target images with an age value greater than the sum of the maximum value and the preset value, that is, 10 images with an age range of The initial target image of 40-50 is removed, and the remaining target image is 80 initial target images in the age range of 30-40 as the remaining target image.

After obtaining the remaining target images, you can calculate the normal age range corresponding to all the remaining target images, for example, you can calculate the mean and standard deviation of the age ranges corresponding to all the remaining target images; obtain the preset hyperparameters; calculate the product of the hyperparameters and the standard deviation , And regard the difference between the mean and the product as the minimum value of the normal age range, and the sum of the mean and the product as the maximum value of the normal age range.

For example, it can be assumed that the result [x1,x2,x3,...xn] predicted by the convolutional neural network model of the image of the same person at any same period conforms to the following Gaussian distribution:

For [x1,x2,x3,...xn], the 4-quartile range method is used to eliminate the obvious outliers and get the remaining m images and their predicted values [x1,x2,x3,...xm ], for the results of these m pictures of the same person, using the Grubbs detection method, calculate the statistical mean u and standard deviation s of the m pictures, design the hyperparameter k, and use the following calculation formula:

μ-k*s≤x _i ≤μ+k*s

For removing the images that are not in the range, the last [x1,x2,x3,...xh] images in the range are obtained, and the average value of the prediction range of these h images is counted as the person's age label As a result, the h infrared images are matched to form a new data set, which is an infrared sample image.

Determining the infrared sample image and the corresponding age annotation result through the convolutional neural network model can save human resources and improve the efficiency of sample acquisition compared with manual age calibration.

S230: Generate a Gaussian distribution of the age annotation result corresponding to each infrared sample image.

For example, for each infrared sample image, a Gaussian distribution with the age annotation result corresponding to the infrared sample image as the center and the preset standard deviation as the peak width can be constructed as the Gaussian distribution of the age annotation result corresponding to the infrared sample image.

The foregoing preset standard deviation may be a preset value, and the embodiment of the present invention does not limit its specific value. It can be understood that the smaller the aforementioned preset standard deviation, the sharper the peak of the generated Gaussian distribution, and the more concentrated the age values included therein.

S240: Input each infrared sample image into the initial infrared convolutional neural network model to obtain the age distribution corresponding to each infrared sample image, and calculate the difference between the age distribution corresponding to each infrared sample image and the Gaussian distribution generated by the corresponding age annotation result, As well as the difference between the expected value of the age distribution corresponding to each infrared sample image and the corresponding age annotation result, the parameters in the initial infrared convolutional neural network model are adjusted according to the calculation results to obtain the infrared convolutional neural network model.

After obtaining the infrared sample image, the corresponding age annotation result, and the Gaussian distribution of the age annotation result, the infrared convolutional neural network model that can predict the age of the infrared image can be trained. Specifically, each infrared sample image can be input into the initial infrared convolutional neural network model to obtain the age distribution corresponding to each infrared sample image, and the difference between the age distribution corresponding to each infrared sample image and the Gaussian distribution generated by the corresponding age annotation result can be calculated And the difference between the expected value of the age distribution corresponding to each infrared sample image and the corresponding age annotation result. According to the calculation result, the parameters in the initial infrared convolutional neural network model are adjusted to obtain the infrared convolutional neural network model.

Specifically, it can construct a distribution learning based on Gaussian distribution estimation and a loss function based on expected age estimation. By converting the face age label into a designed Gaussian distribution as a label, it can be compared with the prediction generated by the model to generate a return error. Adjust the parameters in the initial infrared convolutional neural network model to obtain the infrared convolutional neural network model.

After inputting the initial infrared convolutional neural network model according to each infrared sample image, the difference between the age distribution corresponding to each infrared sample image output by the initial infrared convolutional neural network model and the Gaussian distribution generated by the corresponding age annotation result, and each infrared sample image The difference between the expected value of the corresponding age distribution and the corresponding age labeling result. The parameters in the initial infrared convolutional neural network model are adjusted to obtain the infrared convolutional neural network model. Compared with the scheme that outputs the specific age value, it can be used for the same person The multi-angle and multi-state has accurate age prediction, which improves the robustness of the model.

As an implementation manner of the embodiment of the present invention, in order to ensure the accuracy of the infrared image age prediction, the accuracy detection of the infrared convolutional neural network model obtained by training may be performed.

Specifically, after the infrared convolutional neural network model is obtained, as shown in FIG. 5, the following steps can also be performed.

S510: Obtain an infrared test image and an age annotation result corresponding to each infrared test image; the infrared test image is different from the infrared sample image.

For example, a small amount of infrared images containing human faces can be acquired as infrared test images. In addition, the infrared test images are manually labeled with accurate age.

S520: Determine the test accuracy of the infrared convolutional neural network model according to the infrared test image and the age annotation result corresponding to each infrared test image.

For example, the infrared test image can be input into the infrared convolutional neural network model. After the infrared convolutional neural network model outputs the age distribution of each infrared test image, the age expected value included in the age distribution and the age corresponding to each infrared test image are labeled The results are compared, the accuracy rate is calculated, and it is determined as the test accuracy of the infrared convolutional neural network.

Wherein, when calculating the above accuracy rate, for any infrared test image, the difference between the expected age included in the age distribution and the corresponding age annotation result can be calculated, and the difference is divided by the age annotation result as the error rate. Then calculate the value of 1 minus the error rate as the accuracy rate of the infrared test image. The average value of the accuracy of each infrared test image is used as the test accuracy of the infrared convolutional neural network.

S530: When the test accuracy is less than the preset accuracy threshold, use the current infrared convolutional neural network model as the initial infrared convolutional neural network model, and return to the step of determining each infrared sample image and the age annotation result corresponding to each infrared sample image , Until the test accuracy is not less than the preset accuracy threshold, the current infrared convolutional neural network model is used as the final infrared convolutional neural network model.

When the test accuracy is less than the preset accuracy threshold, it indicates that the age prediction accuracy of the currently trained infrared convolutional neural network model is low. In this case, the infrared convolutional neural network model can be updated to improve its accuracy.

Specifically, the current infrared convolutional neural network model can be used as the initial infrared convolutional neural network model, and the step of determining each infrared sample image and the age annotation result corresponding to each infrared sample image is returned to execute, that is, steps S220-S240. That is to obtain different infrared sample images again, adjust the parameters of the infrared convolutional neural network model, until the test accuracy meets the requirements, use the currently trained infrared convolutional neural network model as the final infrared convolutional neural network model.

Test the test accuracy of the trained infrared convolutional neural network model through the test image, and when the test accuracy is low, update the infrared convolutional neural network model through the infrared sample image again, so as to ensure the final infrared convolution The test accuracy of the neural network improves the accuracy of age prediction.

In an implementation manner, as shown in FIG. 6, the training process of the above-mentioned convolutional neural network model may include the following steps.

S610: Construct an initial convolutional neural network model. The initial convolutional neural network model includes: convolutional layer, pooling layer, and fully connected layer.

The initial convolutional neural network model in the embodiment of the present invention may include data processing layers with parameters, such as a convolution layer, a pooling layer, and a fully connected layer. Among them, the number of convolutional layers, pooling layers, and fully connected layers may be one or more layers, as long as age prediction can be realized, which is not limited in the embodiment of the present invention.

The structure of the initial convolutional neural network model and the foregoing initial infrared convolutional neural network model may be the same or different, which is not limited in the embodiment of the present invention.

S620: Obtain each sample image and the age annotation result corresponding to each sample image.

For example, a color image with an age marked in a public data set can be used as a sample image, and the marked age can be used as the age marking result corresponding to each sample image.

The above-mentioned public data set can be, for example, AFAD: (Asian Face Age Dataset), a public Asian face image data set containing about 160k face images and their age annotations; or, it can be MegaFaceAsia, a public Asian face image data set, contains about 45k face images and their age annotations.

S630: Generate a Gaussian distribution of the age annotation result corresponding to each sample image.

For example, for each sample image, a Gaussian distribution centered on the age annotation result corresponding to the sample image and the preset standard deviation is the peak width can be constructed as the Gaussian distribution of the age annotation result corresponding to the sample image.

S640: Input each sample image into the initial convolutional neural network model to obtain the age distribution corresponding to each sample image, and calculate the difference between the age distribution corresponding to each sample image and the Gaussian distribution generated by the corresponding age annotation result, and each sample image The difference between the expected value of the corresponding age distribution and the corresponding age annotation result, according to the calculation results, adjust the parameters in the initial convolutional neural network model to obtain the candidate neural network model, and adjust the candidate neural network model to obtain the convolutional neural network model.

After obtaining the sample image, the corresponding age annotation result, and the Gaussian distribution of the age annotation result, the convolutional neural network model that can predict the age of the infrared image can be trained. Specifically, each sample image can be input into the initial convolutional neural network model, and the difference between the age distribution corresponding to each sample image output by the initial convolutional neural network model and the Gaussian distribution generated by the corresponding age annotation result, and the corresponding sample image The difference between the expected value of the age distribution and the corresponding age annotation result, the parameters of the initial convolutional neural network model are adjusted to obtain the candidate neural network model, which can predict the age of the color image; then the candidate neural network After the model is adjusted, a convolutional neural network model that can predict the age of infrared images is obtained.

Specifically, it can construct a distribution learning based on Gaussian distribution estimation and a loss function based on expected age estimation. By converting the face age label into a designed Gaussian distribution as a label, it can be compared with the prediction generated by the model to generate a return error. The parameters in the initial convolutional neural network model are adjusted to obtain the candidate convolutional neural network model. Further, the candidate convolutional neural network model is adjusted to a model that can perform age prediction on a single-channel infrared image, and the convolutional neural network model is obtained.

After inputting the initial convolutional neural network model according to each sample image, the difference between the age distribution corresponding to each sample image output by the initial convolutional neural network model and the Gaussian distribution generated by the corresponding age annotation result, and the age distribution corresponding to each sample image The difference between the expected value and the corresponding age annotation result. The parameters in the initial convolutional neural network model are adjusted to obtain the convolutional neural network model. Compared with the solution of outputting specific age values, it can be used for multiple angles and multiple states of the same person. Accurate age prediction improves the robustness of the model.

It can be understood that, in the image to be processed obtained by the electronic device, the human face may face forward, or there may be situations such as a side face other than the front face. When the face is not facing forward, it may affect the accuracy of age prediction results.

As an implementation manner of the embodiment of the present invention, when the electronic device constructs the target image to be processed containing the first face region, it may first perform key point detection on the first face region to obtain the key points of each target in the first face region. The coordinate information of the point; among them, each target key point is a point that identifies the contour feature of the face; then according to the coordinate information of each target key point, the first face area is aligned to obtain the first face area and each The target key point is located in the preset position of the target image to be processed.

For example, based on MTCNN (Multi-task Convolutional Neural Network, multi-task convolutional neural network), the first face area can be detected by key points, and the coordinate information of each target key point in the coordinate system constructed in the image to be processed can be determined , As the coordinate information of each target key point.

In an implementation manner, the above-mentioned key points may include, for example, key points of the eye area. In this way, it is possible to construct a target image to be processed that includes the first face area and the key points of the eye area are located at preset positions.

Perform key point detection on the face area, and then align the face area to obtain the target image to be processed, which can avoid the presence of side faces in the target image to be processed, thereby ensuring that the face in the target image to be processed is clearer and increasing the age The accuracy of the forecast.

As an implementation of the embodiment of the present invention, the electronic device inputs the target image to be processed into the pre-trained infrared convolutional neural network model, and after obtaining the first predicted age distribution of the person corresponding to the first face region, it can also calculate In the first predicted age distribution, the sum of the product of each age value and the corresponding probability, and the calculation result is used as the predicted age value of the person corresponding to the first face area.

According to the predicted age distribution, the specific predicted age value is calculated, so as to obtain the accurate age prediction result.

As shown in FIG. 7, an embodiment of the present invention provides an age prediction device for infrared images, and the device includes:

The infrared image acquisition module 710 is used to acquire an infrared image to be processed;

The face area detection module 720 is configured to detect the first face area in the infrared image to be processed, and construct a target image to be processed that includes the first face area; wherein the size of the target image to be processed Is the default size;

The age prediction module 730 is configured to input the target image to be processed into a pre-trained infrared convolutional neural network model to obtain the first predicted age distribution of the person corresponding to the first face region, wherein the first The predicted age distribution obeys the Gaussian distribution; wherein the infrared convolutional neural network model is inputted into the initial infrared convolutional neural network model according to each infrared sample image, and each infrared sample image output by the initial infrared convolutional neural network model corresponds to The difference between the age distribution and the Gaussian distribution generated by the corresponding age labeling result, and the difference between the expected value of the age distribution corresponding to each infrared sample image and the corresponding age labeling result, adjust each parameter in the initial infrared convolutional neural network model As obtained later, the age distribution obeys the Gaussian distribution; the infrared sample image and the corresponding age annotation result are determined according to a pre-trained convolutional neural network model, and the convolutional neural network model is obtained through color image training.

It can be seen from the foregoing that the device for predicting the age of a face in an infrared image provided by the embodiment of the present invention can determine the infrared sample image and the corresponding age annotation result based on the convolutional neural network model obtained by color image training, and then according to The determined infrared sample images and the corresponding age annotation results are trained to obtain an infrared convolutional neural network model that can predict the age of the face in the infrared image. In addition, compared with manual age calibration, determining infrared sample images and corresponding age annotation results through a convolutional neural network model can save human resources and improve the efficiency of sample acquisition. In addition, when training the infrared convolutional neural network model, after each infrared sample image is input to the initial infrared convolutional neural network model, the initial infrared convolutional neural network model outputs the corresponding age distribution of each infrared sample image and the corresponding age annotation results are generated The difference between the Gaussian distribution of each infrared sample image and the difference between the expected value of the age distribution corresponding to each infrared sample image and the corresponding age annotation result. The parameters in the initial infrared convolutional neural network model are adjusted, which are the same as the output specific age value. In comparison, it can accurately predict the age of the same person from multiple angles and multiple states, and improve the robustness of the model.

Optionally, the device further includes:

Optionally, the infrared sample image determination module includes:

Get preset hyperparameters;

Optionally, the infrared sample image determination module further includes:

Optionally, the device further includes:

Optionally, the face area detection module 720 includes:

Optionally, the device further includes:

The foregoing device embodiment corresponds to the method embodiment, and has the same technical effect as the method embodiment. For specific description, refer to the method embodiment. The device embodiment is obtained based on the method embodiment, and the specific description can be found in the method embodiment part, which will not be repeated here.

A person of ordinary skill in the art can understand that the drawings are only schematic diagrams of an embodiment, and the modules or processes in the drawings are not necessarily necessary for implementing the present invention.

A person of ordinary skill in the art can understand that the modules in the device in the embodiment may be distributed in the device in the embodiment according to the description of the embodiment, or may be located in one or more devices different from this embodiment with corresponding changes. The modules of the above-mentioned embodiments can be combined into one module or further divided into multiple sub-modules.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: The technical solutions recorded in the foregoing embodiments are modified, or some of the technical features are equivalently replaced; these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

An age prediction method for infrared images, characterized in that the method includes:

Obtain infrared images to be processed;

Detecting a first face region in the infrared image to be processed, and constructing a target image to be processed including the first face region; wherein the size of the target image to be processed is a preset size;

Inputting the target image to be processed into a pre-trained infrared convolutional neural network model to obtain a first predicted age distribution of a person corresponding to the first face region, wherein the first predicted age distribution obeys a Gaussian distribution;

Wherein, the infrared convolutional neural network model is based on the input of each infrared sample image to the initial infrared convolutional neural network model, the initial infrared convolutional neural network model outputs the corresponding age distribution of each infrared sample image and the corresponding age annotation result The difference between the generated Gaussian distribution and the difference between the expected value of the age distribution corresponding to each infrared sample image and the corresponding age annotation result are obtained after adjusting each parameter in the initial infrared convolutional neural network model, the age The distribution obeys the Gaussian distribution; the infrared sample image and the corresponding age annotation result are determined according to a pre-trained convolutional neural network model, and the convolutional neural network model is obtained through color image training.
The method according to claim 1, wherein the training process of the infrared convolutional neural network model comprises:

Construct an initial infrared convolutional neural network model, the initial infrared convolutional neural network model includes: a convolutional layer, a pooling layer, and a fully connected layer;

Determine each infrared sample image and the age annotation result corresponding to each infrared sample image;

Generating a Gaussian distribution of the age annotation result corresponding to each infrared sample image;

Input each infrared sample image into the initial infrared convolutional neural network model to obtain the age distribution corresponding to each infrared sample image, and calculate the age distribution corresponding to each infrared sample image and the Gaussian distribution generated by the corresponding age annotation result The difference between the expected value of the age distribution corresponding to each infrared sample image and the difference between the corresponding age annotation results, and the parameters in the initial infrared convolutional neural network model are adjusted according to the calculation results to obtain the infrared convolutional neural network Network model.
The method according to claim 2, wherein the determining each infrared sample image and the age marking result corresponding to each infrared sample image comprises:

Acquiring a plurality of image sets, wherein the initial infrared images in each image set are different facial images of the same person in the same period, and the number of initial infrared images in each image set is greater than a preset number threshold;

For each of the image sets, detecting the second face region in each initial infrared image, and constructing each initial target image including each of the second face regions;

Input each of the initial target images into a pre-trained convolutional neural network model to obtain the second predicted age distribution of the person corresponding to each of the second face regions, and determine the age range corresponding to each second predicted age distribution; Wherein, the convolutional neural network model is based on the input of the initial convolutional neural network model of each sample image, the age distribution corresponding to each sample image output by the initial convolutional neural network model and the Gaussian distribution generated by the corresponding age annotation result The difference, and the difference between the expected value of the age distribution corresponding to each sample image and the corresponding age annotation result, adjust each parameter in the initial convolutional neural network model to obtain a candidate neural network model, and compare the candidate neural network model Obtained after adjustment, the second predicted age distribution obeys a Gaussian distribution, and each sample image is a color image;

For each image set, remove the initial target images with abnormal age ranges in the image set to obtain the remaining target images, calculate the normal age range corresponding to all the remaining target images, and use the remaining target images included in the normal age range as Infrared sample images, and the average value of the age range corresponding to each infrared sample image is used as the age labeling result of each infrared sample image.
The method according to claim 3, wherein, for each image set, removing initial target images with abnormal age ranges in the image set to obtain the remaining target images comprises:

For each image set, sort the initial target images according to the minimum value of the age range corresponding to each initial target image included in the image set in descending order;

Determining a first age range located at one quarter and a second age range located at three quarters, as well as the minimum value of the first age range and the maximum value of the second age range;

Remove the initial target image whose age value is less than the difference between the minimum value and the preset value in the corresponding age range and the initial target image whose age value is greater than the sum of the maximum value and the preset value to obtain the remaining target image .
The method according to claim 3, wherein said calculating the normal age range corresponding to all remaining target images comprises:

Calculate the mean and standard deviation of the age range corresponding to all remaining target images;

Get preset hyperparameters;

Calculate the product of the hyperparameter and the standard deviation, and use the difference between the mean and the product as the minimum value in the normal age range, and use the sum of the mean and the product as the maximum in the normal age range value.
The method according to claim 3, wherein the training process of the convolutional neural network model comprises:

Construct an initial convolutional neural network model, the initial convolutional neural network model including: a convolutional layer, a pooling layer, and a fully connected layer;

Acquiring each sample image and the age annotation result corresponding to each sample image;

Generating a Gaussian distribution of the age annotation result corresponding to each sample image;

Input each sample image into the initial convolutional neural network model to obtain the age distribution corresponding to each sample image, and calculate the difference between the age distribution corresponding to each sample image and the Gaussian distribution generated by the corresponding age annotation result, As well as the difference between the expected value of the age distribution corresponding to each sample image and the corresponding age annotation result, the parameters in the initial convolutional neural network model are adjusted according to the calculation result to obtain the candidate neural network model, and the candidate neural network The model is adjusted to obtain the convolutional neural network model.
The method according to claim 6, wherein said generating the Gaussian distribution of the age labeling result corresponding to each sample image comprises:

For each sample image, construct a Gaussian distribution centered on the age annotation result corresponding to the sample image, and the preset standard deviation is the peak width, as the Gaussian distribution of the age annotation result corresponding to the sample image.
The method according to claim 2, wherein after said obtaining the infrared convolutional neural network model, the method further comprises:

Acquiring an infrared test image and an age marking result corresponding to each of the infrared test images; the infrared test image is different from the infrared sample image;

Determining the test accuracy of the infrared convolutional neural network model according to the infrared test image and the age annotation result corresponding to each of the infrared test images;

When the test accuracy is less than the preset accuracy threshold, use the current infrared convolutional neural network model as the initial infrared convolutional neural network model, and return to execute the determination of each infrared sample image and the age corresponding to each infrared sample image In the step of marking the results, until the test accuracy is not less than the preset accuracy threshold, the current infrared convolutional neural network model is used as the final infrared convolutional neural network model.
The method according to any one of claims 1-8, wherein the constructing the target image to be processed including the first face region comprises:

Performing key point detection on the first face area to obtain coordinate information of each target key point in the first face area; wherein each target key point is a point that identifies a face contour feature;

According to the coordinate information of each target key point, after the first face region is aligned, a target image to be processed including the first face region and each target key point is located at a preset position is obtained.
An age prediction device for infrared images, characterized in that the device comprises:

Infrared image acquisition module for acquiring infrared images to be processed;

The face area detection module is used to detect the first face area in the infrared image to be processed, and construct a target image to be processed containing the first face area; wherein the size of the target image to be processed is Preset size

The age prediction module is used to input the target image to be processed into a pre-trained infrared convolutional neural network model to obtain the first predicted age distribution of the person corresponding to the first face region, wherein the first prediction The age distribution obeys the Gaussian distribution; wherein, the infrared convolutional neural network model is the age corresponding to each infrared sample image output by the initial infrared convolutional neural network model after inputting the initial infrared convolutional neural network model according to each infrared sample image The difference between the distribution and the Gaussian distribution generated by the corresponding age annotation result, and the difference between the expected value of the age distribution corresponding to each infrared sample image and the corresponding age annotation result, after adjusting the parameters in the initial infrared convolutional neural network model It is obtained that the age distribution obeys the Gaussian distribution; the infrared sample image and the corresponding age annotation result are determined according to a pre-trained convolutional neural network model, which is obtained through color image training.