Probabilistic neural networks for segmentation of features in corn kernel images

A method is presented for clustering of pixel color information to segment features within corn kernel images. Features for blue-eye mold, germ damage , sound germ, shadow in sound germ, hard starch, and soft starch were ident ified by red, green, and blue (RGB) pixel value inputs to a probabilistic n eural network. A data grouping method to obtain an exemplar set for adjustm ent of the Probabilistic Neural Network (PNN) weights and optimization of a universal smoothing factor is described. Of the 14,427 available exemplars (RGB pixel values sampled from previously collected images), 778 were used for adjustment of the network weights, 737 were used for optimization of t he PNN smoothing parameter and 12,912 were reserved for network validation. Based on a universal PNN smoothing factor of 0.05, the network was able to provide an overall pixel classification accuracy of 86% on calibration dat a and 75% on unseen data. Much of the misclassification was due to overlap of pixel values among classes. When an additional network layer was added t o combine similar classes (blue-eye mold and germ damage, sound germ and sh adow in sound germ, and hard and soft starch), network results were signifi cantly enhanced so that accuracy on validation data was 94.7%. Image qualit y was shown to be important to the success of this algorithm as lighting an d camera depth of field effects caused artifacts in the segmented images.