<title>Application of normalized gray-scale correlation</title>

Optical pattern recognition based on normalized correlation

“…The shape of the neighborhood is controlled by a window function. Our method guaranties that non-target images never produce a detection signal larger than the noise-free autocorrelation -which could occur with the normalization proposed in [4][5][6][7]. Finally we note, that we are here interested in simple filtering (a single filter corresponds to the detection of one reference target) and that linear or nonlinear filtering combined with the normalization that we analyze, always together define a nonlinear pattern recognition method.…”

“…Due to our interest in efficient optical implementation, here we are only interested in methods which can be expressed in terms of correlations and pointwise processing. The possibility to achieve intensity invariance by a normalization of the correlation signal was recognized by Dickey et all [4][5][6][7]. They proposed to normalize the correlation signal intensity by dividing it by the correlation ofthe squared modulus of the signal and a window function.…”

“…[7] provides intensity invariance with any linear filter h that has a limited impulse response width. In our notation, the normalized correlation signal introduced with use of the Cauchy-Schwarz's inequality in [4][5][6][7] is: c.2 (2) where w is a window that satisfies the conditions: whf = h and (w )2 w . The term s 2 is the correlation of the squared input scene with the window w. Normalization (2) was shown [7] to be practically useful for the intensity invariant pattern recognition problems not only with segmented but also with multiobject input scenes.…”