Fuzzy vector implementation on manifold embedding for head pose estimation with degraded images using fuzzy nearest distance

Advancement of image acquisition and processing technology have triggered the development of 3D face recognition and, along with it, the head poses estimation. The problem arises when image degradation occurred thus reducing the capability of the system to analyze the image. We seek to minimize the problem by constructing a system that handles imprecision data with no significant problem. This paper introduces an alternative approach on manifold embedding head pose estimation on 3D space with 2D intensity image. We employ fuzzy vector used to make the system works with imprecision data thus minimize the negative effect coming from noise and image degradation. On the training set, crisp vector representation of images on specific pose will be transformed to its fuzzy vector representation using a specific triangle fuzzification method. Then, a linear interpolation will be used to construct a manifold, adding data points to improve the precision of pose estimation. In the testing phase, we transform every unknown data image to its fuzzy-vector representation using the parameter we obtained from training phase. We then project the unknown fuzzy vector to the manifolds using a technique called fuzzy nearest distance. The output will be the fuzzy points that mostly represent the unknown fuzzy vector given. This system is applied to recognize pose on images from our database which some of them are influenced by noises. Experimental poses range widely from -90° to 90° horizontally and 0° to 70° vertically. The experimental result shows that the system can correctly recognize horizontal poses with 44.4% success rate and vertical poses with 49.4% success rate.