1 of 12

CNN-Based Age Estimator for Facial Images

Zekun Zhang (112070717)

Xinan Chen (111578939)

12/18/2018

1

2 of 12

Problem and Datasets

  • Biological Age from Facial Image
  • APPA-REAL: 7613 images, rotated & cropped
  • IMDB-WIKI: 50000 images, noisy

2

3 of 12

CNN Architecture

  • Front-End: VGG-16 convolutional layers
  • Back-End: age regressor

3

4 of 12

Training Process

  • 1st Phase: pre-train front-end on ImageNet
  • 2nd Phase: pre-train on IMDB-WIKI (~ 50 hr)
  • 3rd Phase: fine-tune on APPA-REAL (~ 40 hr)

4

2018/12/17 Monday

5 of 12

Best Results (Test Set)

5

Dataset

Testing MAE

A-R

11.99

ImageNet + I-W

7.51

ImageNet + A-R

6.78

ImageNet + I-W + A-R

6.65

Agustsson, et al.

5.47

Agustsson, et al., Apparent and real age estimation in still images with

deep residual regressors on appa-real database. IEEE FG 2017.

6 of 12

Fool the CNN

  • Input Gradient Descent

6

7 of 12

Rule out the Background

  • CNN Focus on Face Region

7

2018/12/17 Monday

8 of 12

Data Augmentation?

  • GAN: Discriminator Improves Generator
  • Here: Generator Improves Discriminator

8

9 of 12

Visualization

  • Manually Modify Image
  • Tell us Where CNN is Looking At
  • Consistent with Common Sense

9

10 of 12

What We Have Done

  • Implemented Age Estimator from Scratch
  • Pre-training Improves Performance
  • CNN can Be Easily Fooled
  • Visualized Where CNN is Looking At

10

11 of 12

11

12 of 12

Attempted Methods

  • AlexNet & ResNet
  • Classifier Network
  • ExprGAN
  • X2Face

12