Enhancing Image Classification with Colorization Techniques and Ensemble Learning

In recent decades, Artificial Intelligence systems have increasingly achieved and surpassed human-level perfor- mance in a variety of complex tasks. Despite their success, the intricate and non-linear structures of deep learn- ing models often make them opaque and challenging to interpret. This thesis presents an innovative automated system for the classification of planktic foraminifera at the species level and extends this methodology to the clas- sification of satellite images from the EuroSAT dataset. The system leverages advanced deep learning techniques, including Generative Adversarial Networks (GANs) and U-Net-based autoencoders. Initially, the foraminifera dataset, comprising 1437 groups of sixteen grayscale images (one group for each specimen), is converted to RGB images through various processing methods. Similarly, the EuroSAT dataset, based on Sentinel-2 satellite images and including 13 spectral bands across 10 classes with a total of 27,000 labeled and geo-referenced images, is also converted to RGB images through diverse processing methods. These newly colored RGB images from both datasets are then classified using transfer learning. The RGB images are fed into a set of Convolutional Neural Networks (CNNs) organized in an Ensemble Learning (EL) environment. The ensemble is built by training different networks using diverse approaches for creating the RGB images, supporting the classifiers to enhance performance. This study demonstrates that an ensemble of CNN models trained on the newly colored RGB images from both datasets improves the system’s performance compared to other state-of-the-art approaches. The main focus of this thesis is to introduce multiple colorization methods that differ from current cutting-edge techniques.