Face Detection In Color Images With Complex Background

Face detection is to locate the faces and determine their sizes in an image. Face detection problem, originated in face recognition research as its first step to locate faces in the images, has recently been intensively researched as an independent problem because of its applications in many different fields including security access control, visual surveillance, content-based information retrieval advanced human and computer interaction.In this paper, the face detection of color still images is discussed. Skin characteristic is used to reduce the search range firstly. Then the facial features extraction comes into being by integrating some information, such as the physical structure features of facial organs, the integral projection, edge images, and the variance projection. The exact face location is detected by the facial features extraction finally.In skin color segmentation, a color balance technique is used to remove the color bias produced by light condition. Meanwhile, a new skin model (H_SI_I skin model) which has the robustness for variety of brightness is built. At last, a cluster-based skin model to remove background points which is wrongly retained by the H_SI_I skin model is proposed.In the facial features extraction, the"holes"are directly matched to locate eyes basing on the physical structure features of facial organs firstly. If eyes can not be gained accurately, facial region's horizontal gray-level projection is used to estimate the horizontal position of eyes, then the edge features are checked by Canny operator, the proper position of eyes is quickly obtained by dealing with"windows"finally. If eyes can not be found still, basing on the horizontal position of eyes, the exact position of eyes is extracted by using the variance projection two times.Experiments show that the proposed face detection algorithm can detect human faces up right or rotated with little angle in color images with complex background, and expression change is tolerated, so the algorithm is feasible and effective.