We present a shape recognition framework which includes two steps: shape searching and shape matching by deformation. First, the user can draw a contour shape descriptor as a search template. The first Bayesian belief propagation (BP I) algorithm is used to find possible targets allowing for translation, scale, and rotation transformations to all contours in a cluttered image. The contour segments with common transformation values are grouped and hypothesized as belonging to the contour in the search template. The search template is then transformed for each possible transformation value. A second belief propagation (BP II) is applied to perform a deformable contour matching. The matching score or cost function determines whether there is an actual match. The algorithm overcomes the weaknesses of the other approaches since it does not require any pre-processing to detect feature points, it can match targets at any position, scale, or rotation transformations, and it does not use any accumulation space that my have peak clustering problems such as in the Hough Transform.
展开▼
