Integrated face detection approach for far image application

• Main Author: Salka, Tanko Daniel
• Format: Thesis
• Language: English
• Published: 2016
• Online Access: http://psasir.upm.edu.my/id/eprint/67088/1/FK%202016%20130%20IR.pdf; http://psasir.upm.edu.my/id/eprint/67088/

Face detection has been widely explored over the past few decades. Despite the significant progress in detecting human faces in unconstrained and complex images, face detection remains a challenging problem in computer vision, especially for the images captured at a far distance making it difficult to detect face region. Other factors affecting face detection are illumination conditions, pose and ethnicity. Therefore, the need of a robust and efficient face detection algorithm is required to tackle these problems. This thesis presents an Integrated face detection approach for far image application, which solves the problems mentioned. The proposed approach consists of an Illumination compensation method, a Skin segmentation method, a Noise reduction method and Euler method. In the proposed illumination compensation method, the R, G and B components were normalized using Gray World Theory (GWT), a theory that compensates the illumination effect. The skin segmentation method consists of a combination of RGB filter, the newly proposed filter known as a Dynamic chrominance filter and an edge detector. The function of the RGB filter is to reject pixels with the RGB colors that are most probably non-skin, so that the computation in the following stages does not apply to all pixels. In this method, the final decision of a pixel belongs to the class “skin is made by the Dynamic chrominance filter and the edge detector. The noise reduction method in the proposed algorithm consists of a combination of a morphological filter and a rejection method. The last stage of the algorithm is to apply the Euler method, in which its function is to search for the facial features. The features indicate whether the detected skin region is a region that represents face or non-face. Also, an experiment was conducted on the developed database known as Large Variability Surveillance Camera Face (LVSC) database and FEI database, the proposed method produced a detection rate of 98.4% and 100%, respectively.