Abstract:

 

Descriptive analysis of consumer products is extensively used to make quality decisions often involving a standard against which test products are compared. Current analysis of this type of data typically involves single attribute tests or even qualitative decisions based on spider plots of the attribute profiles.

​

In 1993, a U.S. patent was issued that described a machine process quite similar to human processing. The patent concerned a process for automatic high speed image inspection of finished product labels in manufacturing. A feature of this patent is that the machine inspection method was based on theory concerning how humans represent percepts in memory and use them in decision making. A machine was presented with multiple typical variants of a package and stored the results in computer memory. Then a new item was presented and the machine’s task was to decide if this item is likely to have been drawn from the same population as those in the training set. Each item in the training set was a photograph containing about 60,000 pixels and these pixels were converted to 64 package segment means with their associated variance-covariance matrix to record the variances of the 64 segments and their dependencies. These data were then used to calculate a statistic leading to accepting or rejecting the item depending on whether it was typical or not of current production.

​

In this technical report we will demonstrate how these ideas can be used to evaluate products based on sensory variables from a descriptive panel.

​