The ideal observer for the task of object localization in the presence of correlated noise is implemented by means of the minimum chi-squared method, which is equivalent to the maximum likelihood method when the noise is additive and normally distributed. The prewhitening approach is explored in which the noise in the data is made uncorrelated by filtering the data with the filter S/sup -1/2/, where S is the noise power spectrum. The location of the object is then found by fitting a model of the object to the prewhitened data. A measure of the goodness of fit is proposed that is based upon the serial correlation in the prewhitened residuals, the difference between the data and the fit. These concepts are demonstrated by applying them to simulated data that possess known noise characteristics. It is shown that the accuracies predicted for the ideal observer can be achieved by the prewhitening technique. (ERA citation 09:025055)
展开▼
