Research On Dynamic Face Detection

Face is one of the most important vision objects in video and digital image and provides a great deal of visual information. Hence, face detection technique has been an important research field in computer vision. Face detection is the first step for any fully automatic face recognition system, and also a step in many surveillance systems towards Automatic Target Recognition (ATR) or generic object detection/recognition.Relying on the analysis of existed face detection techniques, this paper mainly describes face detection based on rectangle feature and dynamic face detection combined with motion information. The main work is as follows:1. An improved AOC algorithm is presented aiming at the inefficiency of original AOC operations. Firstly, the level set of the lowest gray in foreground is extracted, following by AOC operations.Then the level set of next gray level is produced by comparing the result of AOC with the original image. Iteration goes on from low level to high one. Experimental results show that this algorithm can remove lots of small regions early, it improves the efficiency obviously, preserving an equal effect of the original method.2. A face detection algorithm in color image is presented based on feature invariant. Firstly, skin regions are obtained using YCbCr skin color model. Then, eyes and mouth regions are extracted according to their texture features and color features. Finally, a facial geometry structure template is used to validate faces. Experimental results show that this method can fix eyes and mouth exactly.3. A face detection algorithm is implemented based on rectangle features. A fixed size training set is used to train weak classifier for each rectangle feature. The AdaBoost algorithm is applied to promote the performance of weak classifiers to form strong classifiers containing many weak classifiers. Experimental results show that this detector can detect the face fast and exactly, achieving an accuracy of 90%.4. A real-time face detection frame is constructed. At first, skin regions are obtained using a mixture of gaussian skin color model. Then, faces are validated by the face detector based on rectangle features. In order to detect dynamic and static face simultaneously, the results of the front frame are affirmed by difference image to supplement the current frame result. The concept of integral image is introduced to the process of skin color and motion segmentation, which avoids the large amount of computation on the image restoration by traditional methods. Experimental results show this face detection frame satisfies the demand of real-time detection.5. A training data set is constructed, which includes 6000 faces and 18000×20 non-faces, and each face is frontal, equal illumination and uncovered.