A study of the effect of feature reduction via statistically significant pixel selection on fruit object representation, classification, and machine learning prediction

Object recognition or classification has been one of the fundamental foundational building blocks of machine intelligence. Over the years several methodologies have been proposed in the literature. In the past couple of decades, two or three methods have been the predominant means of object recognition; namely Principal Component Analysis, Fisher Linear Discriminant Analysis, and correlation. Considering that a human can easily differentiate between different objects even when the objects are partially obscured, a machine, on the other hand, has greater difficulty in differentiating between objects, even when they are un-obscured. There is important information within a given image that determines the type of object the image contains. This paper presents the usage of a 2-sample statistical t-test as a feature-reduction method to choose those feature pixels of a given image that may be more important and significant than others, and their ordering by order of significance based on a proposed performance criterion metric. The aim is to study the effect of selecting significant feature pixels on the recognition accuracy of the above-mentioned three most popular and widely used object recognition methods. We also introduce a performance criterion that we denote by saturation to evaluate the robustness of the classification/prediction accuracy of these classification methods. We show here that the use of the 2-sample t-test to choose feature pixels and reorganizing these chosen features based upon proposed performance criterion metrics results in many instances in enhancing and stabilizing the recognition results. This paper also introduces for the first time the terms EigenFruit and FisherFruit for eigenvalue based fruit classification and prediction analysis.