On multi-scale feature detection using filter banks

A discrete filtering framework is proposed for multiscale feature detection. The approach starts by choosing the sampling rate for the highest resolution filter of a given signal. Subsequently, a bank of filters for a multiscale representation of that signal is designed by least-square approximation. With this signal representation, multiscale maximum a-posteriori (MAP) detectors are designed for the detection of specific features, such as impulses and edges, without specific knowledge of the signal, the appropriate scale of detection and noise. The approach enables the detection of features with very little or no prior information.