Diversity-based two-phase pruning strategy for maximizing image segmentation generalization with applications in transmission electron microscopy

Abstract For artificial intelligence applications in transmission electron microscopy (TEM), hardware and computational constraints often obstruct real-time data processing, inflating operational costs, consuming valuable instrument time, and heightening the risk of damage to beam-sensitive specimens, thereby complicating reliable data interpretation. To address these issues, we propose a two-stage pruning strategy that reduces deep-learning model size and computational overhead while preserving high performance and generalization across diverse datasets. Unlike conventional pruning techniques, which typically rely solely on weight magnitude and risk overlooking critical variability and directional properties in weight vectors, our approach initially removes filters with low magnitude and insufficient variability, followed by pruning filters with high linear similarity to eliminate redundancy. This one-shot pruning process, followed by fine-tuning, minimizes accuracy loss and mitigates barriers to deep learning integration in TEM workflows. Our method expedites TEM analysis, enabling more efficient, real-time, and cost-effective materials characterization. Additionally, this work lays a foundation for investigating the broader applicability and versatility to different architectures and tasks, particularly in resource-constrained environments where both model size and computational efficiency are critical.