A fast normalizer

Most detection systems use normalizers for dynamic range compression and to provide estimates of signal-to-noise ratio (SNR) from the total power or energy measurements that generally involve signal-plus-noise. Typically, an average of values in a "noise window" of neighboring bins in frequency, time or space is used to obtain a local noise estimate. A critical problem occurs if one or more of the neighboring bins contains a strong signal. Then, the resulting noise estimate will be too high and the SNR estimate too low, causing missed detections. To overcome this problem, a number of computationally intensive approaches have been developed. These attempt to identify and modify or remove from the average those bins that contain the signal. A new approach to this problem is presented, one that has been implemented by a fast convolution and by simple recursion. It involves use of the harmonic mean to suppress strong signals in the "noise window". A statistical analysis is presented. In general, an order-of-magnitude speed-up is achieved over the best normalizer with little or no loss in performance.