A VLSI architecture for computing scale space

Abstract

Meaningful information about a scene is captured in the intensity changes in an image. These intensity changes occur at various scales depending on their physical origins. Scale-space generated by applying the Laplacian of Gaussian edge detector to the image at a continuum of scales is a powerful representation for detecting and organizing these intensity changes symbolically and has proved to be very useful for one-dimensional signals. The high computational cost of generating scale-space in two dimensions has restricted its use in images. This paper proposes a very efficient single chip VLSI design for scale-space computation in one and two dimensions. The architecture of the chip is based on an algorithm that can provide speeds that are of an order of magnitude higher than the speeds obtainable from other systems proposed in the literature. The design uses the principles of modularity, expandability, and parallelism, and fully utilizes the three properties of Gaussian symmetry, separability, and scaling. Moreover, our proposed algorithm does not approximate the Laplacian of the Gaussian operator; it uses instead four one-dimensional convulations to obtain the computations in two dimensions. The proposed architecture has not been built.