Hybrid Optical Neural Network-Type Filters for Multiple Objects Recognition within Cluttered Scenes
Ioannis I. Kypraios · InTech eBooks · 2011
A robust invariant pattern detection and classification system needs to be able to recognise the object under any usual a priori defined distortions such as translation, scaling and inplane and out-of-plane rotation (Wood, 1996) (see Fig. 1). Ideally, the system should be able to recognise (detect and classify) any complex scene of objects even within background clutter noise. This problem is a very complex and difficult one. Here, we will consider only non-deformable (solid) objects (Forsyth & Ponce, 2003). In effect, they maintain their form independent of any of the distortions just described. Early studies (Casasent & Psaltis, 1976) in trying to solve the invariant pattern recognition problem include the system based on a modified logarithmic Mellin transform (Grace & Spann, 1991; Sheng & Arsenault, 1986; Sheng & Lejeune, 1991). Other work (Mersereau & Morris, 1986) has focused on a system based on a circular harmonic filter (Hsu et al., 1982; Hsu & Arsenault, 1982, 1984) illuminated with white light illumination. (Jensen et al., 1987) have described an optical image pattern recognition system based on log-polar mapping (Bryngdahl, 1974; Cederquist & Tai, 1984) of a Fourier transformed input pattern to convert in-plane rotation and scale changes into shift properties. The system’s implementation by a correlator has allowed translation invariance of the input pattern. Although, the real-time practical use of these systems has been superseded, useful concepts for future implementation of filters can be extracted from this work. In literature, broadly, two main categories of pattern recognition systems exist. The first category consists of linear combinatorial-type filters (LCFs) (Stamos, 2001). Proper image analysis in the frequency domain is done with the help of Fourier Transformation (FT) (Lynn & Fuerst, 1998; Proakis & Manolakis, 1988). The second category consists of pure neural modelling approaches. (Wood, 1996) has given a brief but clear review of invariant pattern recognition methods. His survey has divided the methods into two main categories of solving the invariant pattern recognition problem. The first category has two distinct phases of separately calculating the features of the training set pattern to be invariant to certain distortions and then classifying the extracted features. The second one, instead of having two separate phases, has a single phase which parameterises the desired invariances and then adapts them. (Wood, 1996) has also described the integral transforms, which fall