Recognition Of Facing-direction Based On Neural Networks

A general statement of the face recognition problem (in computer vision) can beformulated as follows: Given still or video images of a scene, identify or verify one ormore persons in the scene using a stored database of faces. In most instances the imageswere not taken in a controlled environment. Even the smallest changes in light ororientation could reduce the effectiveness of the system, so they couldn’t be matched to anyface in the database, leading to a high rate of failure.The computer might recognize as different persons by the face images from the sameperson but facing various direction because of the existing large discrepancy betweenfeature vectors corresponding to their direction. To be effective and accurate, the imagecaptured needed to be of a face that was looking almost directly at the camera. With littlevariance of facial orientation from the image in the database leads to quite a problem. So itwould be significant to study the facing-direction recognition which can improve theperformance of face recognition.Taking advantage of the neutral network algorithm to recognize the direction to whichthe person is facing, there exists two group of human faces taken from differentenvironment respectively. The first group consists of50face images from10people with5various orientations, the second consisting of25face images from5people. Then we couldapply them to slove the problem for recognition of facing-direction according to PNN andBP algorithm. And to compare the BP networks with PNN networks under two differentconditions, first we apply both of them to identify the images taken from the sameenvironment and second we recognize the images consists of two different lightbackground.According to the result of experiment, PNN network can work well in the applicationof facing-direction recognition and perform the robust abilities in adapting environment.The accuracy of PNN is100%and perform much better comparing with85%for BP. Onthe other hand, the new examples destroy the accuracy for recognizing original examples in BP network but not occur in PNN network when we mix up the images taken fromdifferent light background.