Multi-face Detection Base Under Complicated Background

As a critical technology of face information processing, face detection has been attracted much attention and become an active research direction in the field of pattern recognition and computer vision application. With the further development of intelligent information processing, face detection are extensively applied in identity recognition, content-based retrieval, automatic surveillance and human-computer interaction.This paper elaborates the face detection method from the perspectives of external structure feature and interior statistical feature. It uses images downloaded from the Internet,CMU and MIT standard face database as the detection objective. We assess the algorithm by its detection result .In the end, the paper combines these two features and applies them into face detection which achieves satisfactory results. Followed is a brief introduction to these methods.In the methods based on external structure feature, this paper mainly adopts the improved Gaussian distribution method and Mid-line location method. These two methods are similar in preprocessing which exclude some candidate areas by facial features and then the number of candidate is confirmed.In the improved Gaussian distribution method, we carefully analyzed relevant papers and tried to carry out using its method after which we found out the detection results unsatisfactory. However, we got expected detection results through reconstructing Gaussian distribution using the image processed by edge detection and eroding. Moreover, the improved method not only realizes more face detection, it also gets better detection rate in a more complicated background with more bare skins in it. The method of mid-line location also used the edge detection image as the processing objective. First of all, the mid-line of the face is located. Secondly, we make the integral round mid-line to get the statistic of " 1" . By using this integral method, the mouth's width and position are found. Finally, this message is used to locate the exact position of the face. The detection results show that it is of fast detection speed and little influence to the changing of expression.Regarding the internal feature method, we adopted adaboost integral method. First, we expended the rectangular features used in Viola by employing the extended rectangular features put forward by Rainer Lienhart. Compared with those rectangular features that Viola used, this rectangular features library add the 45°rotated rectangular features, and it can extend the training range, increase the hit rate, and decrease the false alarm rate. Secondly, we combined the classification of frontal face and profile to face detection. According to this method, we realized face detections of multi-pose. Because we used combined classification, the detection speed is limited. To overcome this defect, the paper integrated external structure feature to exclude candidate area as much as possible and used combined classification to detect face in the candidate area. In the way, detection time is reduced while the detection rate is guaranteed.