Efficient compression of encoder-decoder models for semantic segmentation using the separation index

Abstract

We present a novel approach to compressing encoder–decoder architectures, particularly in semantic segmentation tasks, by leveraging the Separation Index (SI)—a metric that quantifies how distinctly a network’s feature maps separate different classes at the pixel level. By identifying and pruning redundant layers and filters, our method preserves the fine-grained spatial details crucial for segmentation while significantly reducing model complexity. We evaluated our approach on five diverse datasets—CamVid (road scenes), KiTS19 (kidney tumor CT scans), the 2018 Data Science Bowl (nuclei segmentation), Aerial Imagery for remote sensing, and MVTec AD (industrial anomaly detection)—across architectures such as U-Net, LinkNet, MobileNet, DeepLabV3, and SegNet. Experimental results show that SI-driven compression reduces parameters and floating-point operations by up to 70% while maintaining or even improving segmentation accuracy, as measured by mean Intersection over Union (IoU). For example, a compressed DeepLabV3 raises the mean IoU from 0.624 to 0.638 on an aerial imagery dataset with a 2.6× reduction in parameters and faster inference. These findings highlight how SI-based pruning balances efficiency and performance, offering a practical solution for resource-constrained semantic segmentation applications.