Robust machine learning for computer vision in naval application

Abstract

This thesis proposes the development of a resilient machine learning algorithm that can classify naval images for surveillance, search, and detection operations in vast coastal areas. However, real-world datasets may be affected by label noise introduced either through random inaccuracies or deliberate adversarial attacks, both of which can negatively impact the accuracy of machine learning models. Our innovative approach employs Rockafellian Risk Minimization (RRM) to combat label noise contamination. Unlike existing methodologies reliant on extensively cleaned datasets, our two-step process involves adjusting neural network weights and manipulating data point nominal probabilities to isolate potential data corruption effectively. This technique reduces the dependency on meticulous data cleaning, thereby promoting more efficient and timeeffective data processing. To validate the efficacy and reliability of the proposed model, we apply RRM in several parameter configurations to naval environment datasets and assess its classification accuracy against traditional methods. By leveraging the proposed model, we aim to bolster the robustness of ship detection models, paving the way for a novel, reliable tool that could improve automated maritime surveillance systems.