Age and Gender Estimation Using Multiple-Image Features

Abstract

We proposed a real-time system based on multiple frames in this paper to estimate age and gender using facial images. Most of the previous proposed methods are basically based on using a single frame to estimation age and gender. However, limited resources and unpredictable factors on real-time systems are possible to make the result unstable and inaccurate. In order to calibrate the inaccuracy and instability of the previous systems, we decide to construct our system with multiple frames and multiple databases. The first approach we proposed is detecting faces and labeling features from the source images with Stacked Trimmed Active Shape Model (STASM). Then, we perform the alignment of the 76 feature points. Afterwards, we extract features using Speeded Up Robust Features (SURF). After that, we apply the Support Vector Machine (SVM) on the data for preliminary classification. Finally, the data will be sent to the multiple-image and multiple-database classification system to classify the final result. In our experiments, both the training and testing data are from three public available databases which are MORPH, FG-NET, and FERET databases. The experimental result of our proposed method is extremely accurate. Furthermore, the robustness of our system outperforms the previous system based on a single frame.