Face Detection Based On Skin Color And Neural Network

Face detection and recognition technology is an important method for identityauthentication based on biological characters. It has important application in the areas ofsafety access control, video surveillance, intelligent human-computer interfaces andcontent-based image retrieve etc. Aim to construct a usable and practical face detectionsystem, this paper gave a review on the history of face detection at first, then mainlydiscussed the face detection problems in static color image under complex backgroundand varying illumination. It mainly consists of two parts: image segmentingpreprocessing based on skin color and frontal face detection verifying.First, we studied the clustering characteristics of face skin color under some color spaces,and then we choose two of them namely YCbCr and HSI among the models for skinsegmentation according to the contrast between different clustering results. We applied thesegmenting process to an input image under this two color spaces respectively, thencombine the two results into one. This method combined the advantage of the two colormodels, removed some un-skin areas whose color is similar to skin. So it can give moreprecise segmenting result. As a preprocess stage for face detection system, this methodcould quickly remove complex background and put more computing ability on probableface color areas so the executing efficiency and the whole performance is increased. Testsproved that integrating this two color models for skin segmentation could create betterresult that only using one.On the result of skin color segmentation, this paper presents a face detection verifyingmethod using artificial neural network. We trained a BP artificial neural network forfrontal face detection based on the structure that proposed by Rowley etc. As to thesub-image in the face candidate area from the gray image that the color input imageconresponds, we get input window areas in different resolution under fixed scale then inputthem to the face detector. If there is a face, we can get its location and size. The experimentresult shows that this detector achieves a higher executing efficiency and fewer falsedetects than traditional detector.