We present a method for the separation and resynthesis of drum sources from single channel polyphonic mixtures. The frequency domain technique involves identifying the presence of a drum using a novel percussive feature detection function, after which the short-time magnitude spectrum is estimated and scaled according to an estimated time-amplitude function derived from the percussive measure. In addition to producing high quality separation results, the method we describe is also a useful pre-process for drum transcription techniques such as Prior Subspace Analysis in the presence of pitched instruments.