Toward Sustainable Deep Learning: Comparative Analysis and Optimization of Semantic Segmentation Models

begin{abstract}SSemantic segmentation models have become foundational in vision-based applications, yet their environmental impact remains insufficiently explored. In this study, we conduct a comprehensive evaluation of state-of-the-art deep learning architectures for semantic segmentation, jointly analyzing segmentation accuracy, energy consumption, and carbon emissions. We further benchmark the reliability of existing energy tracking tools, identifying the most accurate for measuring training and inference efficiency. Our results reveal a trade-off between performance and environmental cost, with top-performing models often incurring significant energy overhead. To address this, we propose and evaluate targeted optimization strategies, including weight pruning and backbone freezing, that significantly reduce energy usage while preserving segmentation accuracy. Our findings establish a framework for sustainable model development and highlight the importance of integrating environmental efficiency into the design of next-generation computer vision systems.\end{abstract}