In this paper, we propose some ideas to improve the performance of the traditional face detection based on support vector machine (SVM). The traditional SVM-based system for face detection detects faces by exhaustively scanning an image for face-like patterns at any pos More
In this paper, we propose some ideas to improve the performance of the traditional face detection based on support vector machine (SVM). The traditional SVM-based system for face detection detects faces by exhaustively scanning an image for face-like patterns at any possible scales. It divides the original image into overlapping sub-images by using a fixed-size cutting window and classifies them using the Support Vector Machine to determine the appropriate class (face or non-face). This approach has not an acceptable detection rate. In this paper to improve the performance, we use cutting windows with different sizes. We fuse the decisions obtained by using different windows. An important issue in the Support Vector Machine classifier is to shift the decision threshold adequately towards the better represented class. In this paper, a novel method is proposed for determining the threshold value adaptively. A post processing algorithm is also presented for reducing the false alarm rate. Experimental results using standard database show that the performance of the proposed SVM-based method is much better than the basic SVM classifier.
Manuscript profile