A method of age estimation with increased uncertainty prediction

Abstract

The application discloses an age estimation method for increasing uncertainty prediction. The application comprises the following steps: 1, pre-processing and image enhancement are performed on input batch face images; 2, the processed images are input into a feature extraction network for feature extraction, and the extracted feature values are input into an age mean network F and an uncertainty network N for calculation; 3, the age mean network output and the corresponding age uncertainty network output are respectively taken as the mean and variance of a normal distribution, and a predicted age is obtained by sampling from the normal distribution; 4, a loss function is calculated through MSE, and parameters are updated through back propagation; 5, steps 1 to 4 are repeated 50 times for all data in the data set. According to the image enhancement and quantitative introduction of age uncertainty designed according to the face age estimation characteristics, the model has stronger generalization ability, and can more effectively predict the face age.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and more specifically to an age estimation method that increases uncertainty prediction. Background Technology

[0002] Age estimation has wide applications in video surveillance, social networks, and human-computer interaction. Many published methods treat age estimation merely as a precise age calculation problem, thus ignoring the ambiguity and uncertainty inherent in the relationship between facial images and age. Summary of the Invention

[0003] This invention proposes an age estimation method that increases uncertainty prediction, aiming to introduce the ambiguity and uncertainty inherent in the relationship between face images and age into age estimation, so that the age estimation results are more consistent with reality.

[0004] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0005] Step 1: Preprocess and enhance the input batch of face images;

[0006] Step 2: Input the processed image into the feature extraction network for feature extraction, and input the extracted feature values ​​into the age mean network F and the uncertainty network N for calculation;

[0007] Step 3: Using the mean of the age mean network F(x) and the corresponding age uncertainty network output N(x) as the mean and variance of the normal distribution, respectively, sample the predicted age p(x) from the normal distribution;

[0008] Step 4: Calculate the loss function using MSE and update the parameters via backpropagation;

[0009] Step 5: Repeat steps 1 through 4 50 times for all data in the dataset.

[0010] Furthermore, the input image in step 1 is a color original image with RGB three channels.

[0011] Furthermore, in step 1, the preprocessing converts the color originals of inconsistent sizes into images with dimensions of 256*256 using bilinear interpolation, and then performs random edge cropping to crop the images to dimensions of 224*224.

[0012] Furthermore, the image enhancement process includes converting the color image to a black and white image and randomly fine-tuning the brightness of the color image. The final output images to the neural network include the original color image x1, the black and white image x2, and the image x3 after brightness adjustment, all cropped to a size of 224*224.

[0013] Furthermore, the feature extraction network is an improved ResNet18 network, specifically by removing the last fully connected layer of the ResNet18 network.

[0014] Furthermore, step 2 is implemented as follows: Both the age mean network F and the uncertainty network N first pass through three fully connected layers, and the first two fully connected layers have the same structure: 512*256 and 256*128; the last fully connected layer varies depending on the task: the age mean network F uses a classification method to predict age, and the structure of the last layer corresponds to the range of age labels. Let the minimum age be 0 and the maximum age be 69, then the structure of the last layer of F is 128*70. Finally, the index corresponding to the maximum value of the data with dimension 70 is the age F(x) predicted by F; the output N(x) of the uncertainty network N represents the uncertainty of age, and is used as the variance of the output age of F for quantitative calculation. It is predicted using a regression method. Therefore, the structure of the last layer of the uncertainty network N is 128*1, and the output is 1-dimensional, which represents the variance.

[0015] After testing, the beneficial effects of this invention are as follows:

[0016] This invention aims to address the inherent ambiguity and uncertainty between facial images and age in age estimation, making the estimated age more consistent with reality. The key lies in introducing an uncertainty network, which assesses the age uncertainty of the image and ultimately incorporates this uncertainty into the age prediction result when combined with an age mean network.

[0017] Image enhancement designed based on the characteristics of facial age estimation and the quantitative introduction of age uncertainty enhance the model's generalization ability, enabling it to predict facial age more effectively. On the FGNET dataset, which uses only 900 images for training, the average predicted age error was reduced by half a year compared to methods without uncertainty prediction, decreasing from 3.3 years to 2.8 years. Attached Figure Description

[0018] Figure 1 Flowchart of the method of this invention. Detailed Implementation

[0019] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0020] This invention primarily proposes an age estimation method that incorporates uncertainty prediction to improve the accuracy of age estimation in algorithmic models. Figure 1 The specific implementation steps of this invention are as follows:

[0021] Step 1: Batch input of face images, number n, preprocessing and image enhancement. The input images are original color images with RGB three channels; preprocessing converts the inconsistently sized original color images into 256*256 images using bilinear interpolation, and then performs random edge cropping to crop to an image with a length and width of 224*224; based on the fact that black and white images and image brightness have little impact on age judgment, image enhancement includes converting the color images to black and white images and randomly fine-tuning the brightness of the color images. The final output images to the neural network include the original color image x1 cropped to a length and width of 224*224, the black and white image x2, and the image x3 after brightness adjustment.

[0022] Step 2: Input the image x into a feature extraction network (ResNet18 with the last fully connected layer cropped) for feature extraction. The extracted feature values ​​(dimension 512) are then input into two different task networks for calculation (referred to as the age mean network F and the uncertainty network N, respectively). Both F and N first pass through three fully connected layers. The first two layers have the same structure: 512*256 and 256*128. The last layer varies depending on the task. F uses a classification approach to predict age, and the structure of the last layer corresponds to the range of age labels. For example, in FGNET, the minimum age is 0 and the maximum age is 69. The structure of the last layer of F is 128*70. Finally, the index corresponding to the maximum value of the data with dimension 70 is the age F(x) predicted by F. The output N(x) of N represents the uncertainty of age. It is used as the variance of the output age of F in quantitative calculation and is predicted using a regression approach. Therefore, the structure of the last layer of N is 128*1, and the output is 1-dimensional, representing the variance.

[0023] Step 3: Using the output F(x) of the age mean network and the output N(x) of the corresponding age uncertainty network as the mean and variance of the normal distribution, respectively, sample from this normal distribution to obtain the predicted age p(x). The specific technique is to first randomly sample from the standard normal distribution, then multiply by the result of the uncertainty network, and finally add the result of the age mean network to obtain the predicted age.

[0024] Step 4: Calculate the loss function using mean squared error (MSE) and update the parameters via backpropagation.

[0025]

[0026] Where p(x1), p(x2), and p(x3) are the predicted ages of the three images after image enhancement and the above steps, respectively; y represents the real age of the face in the image; and n represents the number of images initially input in batches.

[0027] Step 5: Perform the above steps once for all data in the dataset, repeating 50 times.

[0028] The method obtained through the above steps is based on repeated testing and can effectively improve the model's age estimation ability.

Claims

1. An age estimation method that increases uncertainty prediction, characterized in that... The steps include the following: Step 1: Preprocess and enhance the input batch of face images; Step 2: Input the processed image into the feature extraction network for feature extraction, and input the extracted feature values ​​into the age mean network F and the uncertainty network N for calculation; Step 3: Using the mean of the age mean network F(x) and the corresponding age uncertainty network output N(x) as the mean and variance of the normal distribution, respectively, sample the predicted age p(x) from the normal distribution; Step 4: Calculate the loss function using MSE and update the parameters via backpropagation; Step 5: Repeat steps 1 to 4 50 times for all data in the dataset; Step 2 is implemented as follows: Both the age mean network F and the uncertainty network N first pass through three fully connected layers, and the first two fully connected layers have the same structure: 512*256 and 256*128. The last fully connected layer varies depending on the task: the age mean network F uses a classification method to predict age, and the structure of the last layer corresponds to the range of age labels. Let the minimum age be 0 and the maximum age be 69, then the structure of the last layer of F is 128*70. Finally, the index corresponding to the maximum value of the data with dimension 70 is the age F(x) predicted by F. The output N(x) of the uncertainty network N represents the uncertainty of age, and it is used as the variance of the output age of F for quantitative calculation. It is predicted using a regression method. Therefore, the structure of the last layer of the uncertainty network N is 128*1, and the output is 1-dimensional, which represents the variance. Step 3 involves first randomly sampling from the standard normal distribution, then multiplying by the result of the uncertainty network, and finally adding the result of the age mean network to obtain the predicted age.

2. The age estimation method with increased uncertainty prediction according to claim 1, characterized in that... The input image in step 1 is a color original image with RGB three channels.

3. The age estimation method with increased uncertainty prediction according to claim 2, characterized in that... In step 1, the preprocessing converts the original color images of inconsistent sizes into images with dimensions of 256*256 using bilinear interpolation, and then performs random edge cropping to crop the images to dimensions of 224*224.

4. An age estimation method with increased uncertainty prediction according to claim 2 or 3, characterized in that... The image enhancement includes converting a color image to a black and white image and randomly fine-tuning the brightness of the color image. The final output image to the neural network includes the original color image x1, the black and white image x2, and the image x3 after brightness adjustment, all cropped to a size of 224*224.

5. The age estimation method with increased uncertainty prediction according to claim 4, characterized in that... The feature extraction network is an improved version of the ResNet18 network, specifically by removing the last fully connected layer of the ResNet18 network.

6. The age estimation method with increased uncertainty prediction according to claim 5, characterized in that... Step 4 loss function calculation is as follows: ； Where p(x1), p(x2), and p(x3) are the predicted ages of the three images after image enhancement and the above steps, respectively, y represents the real age of the face in the image, and n represents the number of images initially input in batches.

Images