Feasibility of Deep-Learning Methods for Automatic Fruit Classification Tasks

Abstract

The adoption of Machine Learning techniques has steadily increased over the last years. The most successful results have been established in the Deep Learning field and in particular using the Convolutional Neural Networks (CNNs) framework. These kind of architectures are able to achieve classification results exceedingly close to human judgment, by leveraging on very large training sets of labeled images. In this thesis we study how it is possible to apply this approach to solve a specific industrial problem, that is the fruit classification. In particular we compare the performance obtained by means of different architectures, ranging from the current state-of-art to specially crafted networks. To this end, we explore a lot of different aspects, including the depth of the network, the type and the sequence of the layers, the dimension of the kernels, the preprocessing of the images and we specifically address the underfitting or the overfitting problems. The evaluation is focused on the recognition of production quality for different breeds of olives, by learning the features of both good and not-for-sale fruits.