Automatic corn (Zea mays) kernel inspection system using novelty detection based on principal component analysis

Corn (Zea mays) kernel processing companies evaluate the quality of kernels to determine the price of a batch. Human inspectors in labs inspect a reduced set of kernels to estimate the proportion of damaged kernels in any given lot. The visual differences between good and damaged kernels may be minor and, therefore, difficult to discern. Our goal is to design a computer vision system that enables the automatic evaluation of the quality of corn lots. To decide if an individual kernel can be accepted or rejected, it is necessary to design a method to detect defects, as well as quantify the defective proportions. A setup to work in-line and an approach to identify damaged kernels that combines algorithm-based computer vision techniques of novelty detection and principal component analysis (PCA) is presented. Experiments were carried out in three colour spaces using 450 dent corn kernels previously classified by experts. Results show that the method is promising (92% success) but extensions are recommended to further improve results.