Pattern Recognition in Time-Frequency Domain: Selective Regional Correlation and Its Applications

Pattern recognition is a very powerful tool in automated data analysis and it is widely used in many different applications (Chou & Juang, 2003; Jiang,1994; Blue et al., 1994; Milosavljevic, 1994; Moreels & Smrekar, 2003). However, the application of such a tool can be a difficult task in some cases. For example, in a correlation-type scheme, the basic idea is to correlate the signal being analyzed with a known template or templates (Shiavi, 1999; Scharf, 1991) and make decisions based on the magnitude of the correlation coefficients, which is between 0 and 1. In practice, these extreme values are seldom achieved due to corrupting signals/noise that can affect the accuracy of pattern matching and subsequently lead to errors in classification (Kil & Shin, 1996). The corrupting signals may also bear some resemblance to the template being matched. This is particularly true if the pattern of interest is a non-stationary transient signal. Furthermore, it is well known that traditional time domain correlation-based pattern recognition methods do not fully utilize the frequency characteristics of the template and the signal being analyzed. Hence, such methods perform poorly when applied to transient signals. To overcome these difficulties, a scheme known as selective regional correlation (SRC) has been developed (Sejdic & Jiang, 2007). It has been shown that if a template has bandlimited characteristics, significant improvement in the performance of pattern recognition can be readily made by a relatively simple preprocessing of the signal and the template in the time-frequency domain (Sejdic & Jiang, 2007). The redundant representation of a 1D signal in a 2D time-frequency domain can provide an additional degree of freedom for signal analysis. Such pre-processing effectively separates the intertwined time domain features of the signal, allowing the important characteristics to be exposed in the time-frequency domain, resulting in more effective pattern matching. Hence, correlation between the signal being analyzed and the template needs to be conducted only in selected regions of interest in the time-frequency domain. An overview of the theoretical developments behind the SRC is provided in this chapter along with some recent results. The performance of the scheme is briefly reviewed and compared with that of the general correlation technique through the analysis of a set of O pe n A cc es s D at ab as e w w w .ite ch on lin e. co m