Feature selection for a real-time vision-based food inspection system

In a world where automation of processes is more and more on demand, machine vision is continuously explored to address several industrial problems such as quality inspection. In the processed-food industry where the external quality attributes of the product are inspected visually before the packaging line, machine vision systems often involve the extraction of a larger number of features than those actually needed to ensure proper quality control. This work experiments with several feature selection techniques in order to reduce the number of attributes analyzed by a real-time vision-based food inspection system. Four filter-based and wrapper-based feature selectors are evaluated on seeded buns and tortillas datasets. Experimental results show that consistency-based and the RELIEF subset evaluation techniques perform the best for all the considered datasets in terms of accuracy. However, variations in the number of attributes selected still vary significantly between these techniques.