Many current imaging techniques produce multi-channel images. For example ultrasonic test data typically consists of reflection energy at multiple times, as the projected test pulse is reflected from front and rear surfaces, as well as internal features. Each pixel on the ultrasonic image will therefore have possibly hundreds of channels. Practical methods of decoding the images often use thresholds set on an empirical basis. The aim of this work is to provide a framework for extracting information from these multi-channel images based on physics and mathematics and applicable to a wide range of imaging techniques.

Principle Value Decomposition (PVD) can be used to significantly reduce the volume of data in multi-channel images. The images below show transient eddy-current data acquired from a multilayer aluminium panel. The acquired data consists of 18 channels; each channel represents a time-slice through the structure. Using PVD the channels are decomposed so that 98% of the total image intensity variance is contained in just 3 channels. The remaining channels consist mainly of noise. 

Left: Before PVD, showing one channel of eighteen.  Middle: After PVD, channel containing most variance. Right: After PVD, channel containing least variance (mostly noise).