Vision system development by machine learning: Mashing assessment in brewing

This article describes the development of a machine vision application for automatic process assessment by image analysis and machine learning. The system is required to differentiate between the various stages of a mashing process and determine the termination point. A large number of histograms, Haralick and Gabor features (835) were extracted from 275 training images. Three feature selection algorithms - wrapper, consistency filter, and correlation filter - were then applied to the training data, resulting in feature sets of size 29, 15, and 11, respectively. A number of decision tree, rule induction, and nearest neighbor classification algorithms were then applied to the reduced data set. For discriminating seven stages of the mashing process, the highest accuracy obtained was 71.6%. For the binary problem of differentiating the finished state from all of the other states the accuracy was 92.0%. This accuracy is good enough for deployment. The results indicate that using a large library of features and machine-learning methods for removing redundant features can significantly reduce development times for vision systems by eliminating the time-consuming manual search for the best discriminating features.