An improved OTSU based watershed segmentation and its implementation on Virtex II pro platform

Watershed algorithm as was proposed by Vincent and Soille is an elegant formalism for image segmentation. Watershed conventional algorithm finds out peaks in the gradient image and identifies them as contours. Although this technique has proven its efficiency in some previous works, it suffers from over-segmentation problems, due to the fact that it considers in a similar way lower and higher valued gradients. This paper overcomes watershed over-segmentation problems by introducing a pre processing step based on OTSU thresholding method, in which a pre selection of the higher gradients will keep only real contours in the image. Moreover, an anisotropic diffusion is applied for smoothing the gradient image, getting rid of noisy peaks. The whole design was implemented on a Virtex II pro platform. An optimization of the hardware/Software resources was done based on a co-design methodology. On a V2pro ff672, the design was applied to a 128*128 pixel image, leading to a 116MHz frequency. Performance evaluation was done on some images taken from different classes of the Berkley database.