I want to build a classifier for sounds, but I’d like to pre-process the images to remove unreliable structure. Is this possible?

Answer:

Time-frequency microstructure is unstable for regions of the time-frequency plane with spectrally dense content. For those regions, small changes in analysis parameters or added background noise in the signal can lead to changes in the details of a sonogram or contour shapes. Structurally unstable portions of the representation can be eliminated by showing only contour fragments that are in agreement across different angles and time-scales of analysis.

The image below illustrates that process. On top, all long contours are shown, weighted by sonogram power.
On bottom, only the structurally stable “consensus” elements are shown, also weighted by sonogram power.