| Research on face is of great importance for its representative as a kind of characteristic in human identification.Face detection is defined as a process that finding out the number,location and scale of the possible faces in the test images.As it can conduct by non-contact way,face detection has broad application prospects in the public security,finance,aviation,customs and even national and military security and other video surveillance.Early face detection studies often assume that human face location is easy to obtain,mainly to face images with strong constraints.However,in practical application,face detection is not only confronted with the change of postures and details about facial features,but also changes of illumination,occlusion of the face and change of imaging conditions.These changes increase the difficulty of optimizing the performance of face detection methods,and so,face detection under unconstrained conditions has attracted attention of researchers.The visual attention can help the brain filter out a large amount of interference information from unconstrained conditions,so that attention can focus on the subjects which one concerned and eyes can deal with the visual information effectively.Yet,the face salient features can not be well represented by the existing visual saliency models,further research is need to find the ways in which visual saliency can be applied to face detection.In this paper,a method using calculation model of face saliency based on part-based model and the theory of saliency is proposed for the problem of face detection in unconstrained conditions.For each test image,Face saliency maps under three perspectives out of planeare generated,then a complete and effective face saliency map can be obtained.The results shows that the face saliency map which we obtained can describe the region of face accurately,and our face detection method honors the better performance compared to the original part-based model.The VJ face detector with the feature of Haar can get good results for images which contain almost front faces,while it is not suitable for the faces which is out of plane.The face detection method based on part-based model with feature of HOG can perform well for the multi-pose face while it takes a long time.In this paper,the advantages of these two methods are combined,what’s more,the theory of saliency is introduced so that the face saliency maps of these two salient features of Haar and HOG can be generated,the fusion of these two saliency maps of face help detecting face in unconstrained conditions.The face detection method based on HPM(Hierarchical Deformable Part Model),which is suitable for detection of occluded faces can model both face and the occlusion pattern.In this paper,experiments and research are conducted.Results shows that the chosen of-90 degrees,-45 degrees,0 degrees,45 degrees,and 90 degrees as the model angles can enhance the performance and Robustness for the face detection in unconstrained conditions. |
