Shibei Lu MSc Dissertation 2015/16

Improved Multi-side Face Detection

Supervised by Y.Gotoh

Abstract

The final goal of this project is to develop a program to detect face and estimate the direction that the person is facing with still picture. The research part is consist of face detection algorithms and facial feature algorithms. Finally, the combination of two techniques will create an improved face detection measurement using MATLAB.

The combination of two algorithm is applied to the program. A set of pictures is used to test the program. The dataset is pictures including one person in many different conditions such as brightness, gender. The program tests each pictures to determine the direction of the person is facing to. There are 4 types: frontal face, face at right 45, 90, and 135 degrees.

The success rate of the testing results for each type are respectively 86.67%, 73.33%, 75% and 70%. By comparing and testing, it is found that the final result can be influenced by many factors, such as over darkness, wearing glasses and same skin-color objects, which is mainly due to the lack of robustness of skin-color algorithm and the lack of training samples of HOG feature.

Overall, this project is properly completed. Although the program still need to be improved by some future work, it is already developed successfully to detect the face region and determine four directions of the face.