2012, Saarland University, Germany.
An age group classification system has been developed, and the implementations using 3 different, yet powerful vision approaches arecompared. Three age groups, including babies(0-20), middle-aged adults(20-50)and old-aged adults(>50) , are the categories of classification system. The process of the system is divided into 3 major phases: patch extraction, feature extraction and age classification. Interest patches are extracted. Sobel edgeoperator, geometric distance ratios and local binary patterns are used to extractthe features out of interest patches. Two SVM models are trained for classification. The first one employs the sobel edges and geometric ratio features to distinguish whether the facial image is baby. If it is not, then the second model uses Local binary patterns in addition to sobel and geometric features, to classify the image into one of the two adult groups. The proposed system is experimented with 150 images. 100 images are used for training purposes and 50 are used for testing purposes. The system has been developed using Matlab toolkit.
1. The sobel edge method for classification into 2 age categories, using thresholding parameter : 4. Accuracy - around 60%
2. The LBP along with SVM, for classification into 2 age categories, using‘C’ parameter: 2 (approx). Accuracy: 60 % - 70 % where Train Images: 100, Test Images: 50.
The accuracy varies heavily with the choice and size of training and test data and the value of ‘C’ parameter.
3. The sobel edges, geometric ratios and normalised LBP features along with SVM, for classification into 3 age categories, using thresholding parameter: 4
‘C1’ parameter for SVM model 1 : 1e3
‘C2’ parameter for SVM model 2 : 1e2
Accuracy: 64 % - 70 % where,Train Images: 50/100Test Images: 50.
The accuracy varies heavily with the choice and size of training and test dataand the value of ‘C’ parameter.