Detection And Analysis Of Face Occlusion

Partially occluded face is the face blocked by sunglasses, scarf, masks, helmets and other obstructions. As a result, the face cannot be tested or classified by face detection system. The purpose of facial occlusion detection and analysis is to estimate the occlusion possibility of the face image by analyzing and processing the acquired images.This thesis proposed an unsupervised-learning algorithm based on FSDP (Fast Search and Find of Density Peaks) to solve the problem of facial occlusion detection. The position of head area was extracted in video sequences and it was detected by facial occlusion detection system. If the sample is occluded face, the five-organ was detected by this system.In the section of head area location, this thesis designed a method, based on moving target detection and SVM algorithm. Gaussian mixture background model of moving target detection algorithm was used to extract moving targets, and the area of moving object was determined by the pixel density from the horizontal and vertical direction in a video image. The SVM classifier located the head area effectively with the input of moving area in the video. And then, the fuzzy positioning of five-organ extraction method, based on Active Shape Models was studied with the help of stasm open library.In the part of occlusion detection, this thesis studied the facial occlusion detection algorithm based on FSDP algorithm with different feature extraction methods and different scales of sample. At last, this paper proposed a new online learning algorithm based on FSDP.The experiment results show that the unsupervised-learning method avoided the complexity of the sample labeling task and this method can achieve a good face occlusion decision in a small number of samples and has a very good robustness of the illumination and gender of the images.