Robust and Computationally Efficient Face Detection Using Gaussian Derivative Features of Higher Orders

In this paper, we show that a cascade of classifiers using Gaussian derivatives features up to fourth order can be used efficiently to improve the detection performance and robustness as well when compared with the popular approaches using Haar-like features or using Gaussian derivatives of lower order. We also present a new training method that structures the cascade detection so as to use the least expensive derivatives in the initial stages, so as to reduce the overall computational cost of detection. We demonstrate these improvements with experiments using two publicly available datasets (MIT+CMU and FDDB), in the face detection problem, in addition we perform several experiment to show the robustness of Gaussian derivatives when several transformations are presented in the image.