Programming techniques for real time software implementation of optimal edge detectors :-a comparison between state of the Art DSPs and RISC architectures -

This paper presents the real time implementations of the Canny-Deriche optimal edge detectors on RISC and DSP processors. For each type of architecture, the most leading optimization techniques are described. A comparison is then made between DSP and RISC processing speeds. INTRODUCTION Canny-Deriche operators have asserted themselves to the edge detection field which stands as a fundamental component of image processing and computer vision. The main drawback is the prohibitive computational power they require. It has led people to design dedicated hardware implementations (FPGA, ASIC) to achieve the real time execution of these detectors. On the other hand, the still increasing performance of DSP and RISC calls into question the need for a dedicated architecture. This paper shows that crafty software implementation of Canny-Deriche edge detectors may achieve real time execution on state of the art DSP and RISC processors. To achieve this goal, we introduce data parallelism techniques as well as flow optimization methods. 1 OPTIMAL EDGE DETECTORS In this section we introduce optimal edge detectors and the derived filters we will use in the subsequent sections. 1.1 Canny’s optimal filters Canny’s approach [1] consists in finding the optimal FIR filter which satisfies the three following constraints for a Heaviside input signal: good detection, good localization, low maximum multiplicity due to the noise. Deriche [3], using Canny’s approach, has looked for an IIR filter which satisfies the same constraints. He got the same differential equation, but while changing the conditions at the limits, he obtained, for the Canny’s performance index, an improvement of 25%. Deriche’s operators are used for two important methods of edge detection. The first is based on the gradient maxima, the second, on the laplacian's zero crossings. The state of the art methods combine both of them. In this paper we will focus on the gradient method. The implemented operators are those proposed by Deriche in [3]. 1.2 Deriche’s gradient The horizontal derivative is the result of a smoothing in the vertical direction followed by a derivation in the horizontal direction. Respectively, the vertical derivative is based on those transposed directions. The smoothing operator is the sum of a causal and an anti-causal filter: [ ] ) 2 ( ) 1 ( 2 ) 1 ( ) 1 ( ) ( ) ( 1 2 1 1 − − − + − − + = − − − n y e n y e n x e n x k n y α α α α [ ] ) 2 ( ) 1 ( 2 ) 2 ( ) 1 ( ) 1 ( ) 1 ( ) ( 2 2 2 2 + − + + + − − + − = − − − − n y e n y e n x e n x e k n y α α α α α α ) ( ) ( ) ( 2 1 n y n y n y + = , α α α