Face detection for TV video using cascaded decision trees and Gabor convolution

Abstract

An effective technique of retrieving desired video scenes is necessary to manage huge video archive. Especially at the broadcasting station, there are demands of producers to retrieve video scenes with specific person. Firstly, we need to detect face position at a video frame before recognizing person. But existing face detection methods have an issue that the accuracy decreases when target video includes large variation of appearance such as illumination condition, facial direction and facial expression. This paper proposes a novel face detection method for TV video which is robust to such video variation. The proposed method uses cascaded decision trees and Gabor convolution filter to improve both detection accuracy and processing cost. In experiment, we used broadcast TV video to verify an effectiveness of the proposed method. We also performed comparison experiment with several existing methods to verify superiority of our method.