Diversity-Based Two-Phase Pruning Strategy for Maximizing Image Segmentation Generalization with applications in Transmission Electron Microscopy
Abstract
To address the storage and computational demands of Transmission Electron Microscopy (TEM), we propose a two-phase pruning strategy that reduces model size and enhances speed while maintaining performance across diverse datasets, practical for TEM analysis. Unlike traditional pruning methods that focus solely on weight magnitude, our approach also considers weight variability to preserve feature diversity, crucial for generalization in the varied context of TEM images. Our strategy first prunes filters with low magnitude and variability, then removes redundant filters with high linear similarity. This two-phase pruning, followed by fine-tuning, effectively reduces parameters and computational load while ensuring high accuracy and generalizability.