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

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Encoder-decoder architectures are vital for semantic segmentation but often suffer from high computational complexity.
• Preserving fine-grained spatial details is crucial for accurate segmentation, posing a challenge for model compression.

Purpose of the Study:

• To develop and evaluate a novel compression approach for encoder-decoder networks in semantic segmentation.
• To leverage the Separation Index (SI) for identifying and pruning redundant network components.

Main Methods:

• Utilized the Separation Index (SI) to quantify feature map distinctiveness for class separation at the pixel level.
• Implemented a pruning strategy targeting redundant layers and filters within various semantic segmentation architectures.
• Evaluated the approach on diverse datasets (CamVid, KiTS19, Data Science Bowl, Aerial Imagery, MVTec AD) and architectures (U-Net, LinkNet, MobileNet, DeepLabV3, SegNet).

Main Results:

• Achieved significant reductions in model parameters and floating-point operations (up to 70%) through SI-driven compression.
• Maintained or improved segmentation accuracy, measured by mean Intersection over Union (IoU), across all tested datasets and architectures.
• Demonstrated a compressed DeepLabV3 model improving mean IoU from 0.624 to 0.638 on aerial imagery with a 2.6x parameter reduction and faster inference.

Conclusions:

• SI-based pruning offers an effective method to balance model efficiency and performance in semantic segmentation.
• The proposed approach provides a practical solution for deploying deep learning models in resource-constrained environments.
• This work highlights the potential of feature separability metrics for guiding efficient deep learning model compression.