Age recoginition via fastAAMs

A very simple approach!

feature extraction

The code is in the featureExtraction.m file.

It includes two steps:

train AAM Model

featureExtraction.m

Set

% featureExtraction.m

% the training data is in `morph_small/trainset/`

flag_train = 1;
flag_precalibration == 1

Run featureExtraction.m script,

• Set flag_train = 1;, then AAM.mat file (AAM model) will be generated in the folder morph_small/trainset/. After that, set flag_train = 0;.
• Set flag_precalibration == 1;, then the calibration parameters would be computed and saved in meanscl.mat and meantrans.mat. After that, set flag_precalibration = 0;.

extract the training images using the AAM model

The training/test images are in the folder morph_small/Images_ori/ (total 2500 images).

dataset the morph dataset can be download from this website.

Alternatively, I upload some images here (total 10000 images) and here (total 28533 images) for academic use.

After run featureExtraction.m,

• the piecewirse affine warpped images are generated in the folder morph_small/Images_normalized/,
• the images with extracted features are generated in the folder morph_small/Images_withfeatures/,
• and the extracted batch features in the .mat form are saved in the folder morph_small/features_mat0/.

collect all extracted features in a single .mat file

append_features.m

If the images dataset is very large, it may take a long time to train. So you can partition the datasets into multiple parts, and train each part separately. Just remember rename the destination folder, like features_mat0, features_mat1, features_mat2, features_mat3.

Then run the append_features.m script.

This script will collect all features from batch features into a single .mat file, Allfeatures.mat.

age classification

main.m

After run featureExtraction.m and append_features.m, the features are generated in Allfeatures.mat.

Then run main.m.

The script main.m train and test a classification model using Allfeatures.mat.

Set is_partition_dataset == 1 when first runs to generate trainset.mat and testset.mat.

Set is_svmtrain == 1 and set svm_type= 0 or 1 or 2 to train a svr model.

Set is_svmtest == 1 to test the model.

test images

test_images.m script predicts the age from images.

Report

REPORT is given: age_estimation_report.pdf

Reference

[1] Xin Geng, Zhi-Hua Zhou, and Kate Smith-Miles. Automatic age estimation based on facial aging patterns. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 29(12):2234–2240,2007.

[2] Georgios Tzimiropoulos and Maja Pantic. Optimization problems for fast aam fitting in-thewild. In Computer Vision (ICCV), 2013 IEEE International Conference on, pages 593–600.IEEE, 2013. http://www.mathworks.com/matlabcentral/fileexchange/44651-active-appearance-models--aams-

[3] Khoa Luu, Karl Ricanek, Tien D Bui, and Ching Y Suen. Age estimation using active appearance models and support vector machine regression. In Biometrics: Theory, Applications, and Systems, 2009. BTAS’09. IEEE 3rd International Conference on, pages 1–5.IEEE, 2009.

[4] Ricanek, Karl, and Tamirat Tesafaye. "Morph: A longitudinal image database of normal adult age-progression." Automatic Face and Gesture Recognition, 2006. FGR 2006. 7th International Conference on. IEEE, 2006.

---

Contact: huajh7@gmail.com

2014/6/19