Design of High Speed Neural Networks for Fast Pattern Detection by using Cross Correlation and Matrix Decomposition

Abstract

Fast pattern detection and identification is a fundamental problem for many applications of real-time systems (Bruce & Veloso 2003). Its reliability and performance have a major influence in a whole pattern recognition system. Nowadays, neural networks have shown very good results for detecting a certain pattern in a given image (Rowley et al. 1998; Feraud et al. 2000; Anifantis et al. 1999; Lang et al. 1988; El-Bakry 2001). Among other techniques (Schneiderman & Kanade 1998; Zhu et al. 2000; Srisuk & Kurutach 2002; Bao et al. 2006), neural networks are efficient pattern detectors (Rowley et al. 1998; Feraud et al. 2000; ElBakry 2002,a; El-bakry 2002,b; Essannouni and Ibn Elhaj 2006; Roth et al. 2006; Ramasubramanian & Kannan 2006). But the problem with neural networks is that the computational complexity is very high because the networks have to process many small local windows in the images (Zhu et al. 2000; Srisuk & Kurutach 2002; Yang et al. 2002). The main objective of this paper is to reduce the detection time using neural networks. The idea is to accelerate the operation of neural networks by performing the testing process in the frequency domain instead of spatial domain. Then, cross-correlation between the input image and the weights of neural networks is performed in the frequency domain. This model is called fast neural networks. Compared to conventional neural networks, fast neural networks show a significant reduction in the number of computation steps required to detect a certain pattern in a given image under test. Furthermore, another idea to increase the speed of these fast neural networks through image decomposition is presented. Moreover, the problem of sub-image (local) normalization in the Fourier space which presented in (Feraud et al. 2000) is solved.. The number of computation steps required for weight normalization is proved to be less than that needed for image normalization. Also, the effect of weight normalization on the speed up ratio is theoretically and practically discussed. Mathematical calculations prove that the new idea of weight normalization, instead of image normalization, provides good results and increases the speed up ratio. This is because weight normalization requires fewer computation steps than sub-image normalization. Moreover, for neural networks, normalization of weights can be easily done off line before starting the search process. In section 2, high speed neural networks for pattern detection are described. The details of conventional neural networks, high speed neural networks, and the speed up ratio of