Abnormal Face Detection Based On And-Or Graphs

With the development of information technology and networking, Face Detection technology has been widely used. The basic idea of face detection is based on knowledge or statistical modeling of the face, matched with a detection area of all possible models of human face, finding the face region. Currently there are many FD algorithms, based on the features adopted, can be divided into five models: color-based model, geometry-based model, prior-knowledge model, statistics-based model and information associated model. Actually, the face is often obscured, wearing sunglasses, wearing a face mask and face with a lower resolution, which we called Abnormal Face,and there is not a universal robust FD method can deal with these situations . In view of this situation, we proposed a multi-channel method based on And-Or Graph for Abnormal FD. Three channels are defined:α-channel,β-channel,γ-channel.α-channel detect face directly,β-channel get face region by detecting the local features of the face,γ-channel get the face area by detecting the head and shoulder area.α- channel combines color model and AdaBoost method for face detection, which can improve the efficiency and robustness.β- channel first usees local feature detection method to detect the eyes, nose, mouth, and then integrates the detection results by bottom-up method to get face region.γ- channel uses HOG features and SVM classifier to detect head and shoulders, and speculates face by top-down method. And-Or Graph is used to express the three processes, and we set the weights for each channel, then we use the greedy way to integrate the results of three channels to get the final result. By experiments, whenα-channel detection of human faces failed, we still can determine the face byβorγ-channel,which improves the robustness of face detection. Multi-channel FD technology can be a good choice to deal with obscured face, etc., and realize the Abnormal FD.