Lightweight Class Incremental Semantic Segmentation Without Catastrophic Forgetting.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Class Incremental Semantic Segmentation (CISS) models typically require significant computational resources, hindering their use on edge devices.
• Existing CISS methods often struggle with catastrophic forgetting when model parameters are reduced.

Purpose of the Study:

• To develop a Lightweight Class Incremental Semantic Segmentation (LISS) model for resource-constrained environments.
• To improve the efficiency and effectiveness of continual learning in semantic segmentation tasks.

Main Methods:

• An automatic knowledge-preservation pruning strategy using Hilbert-Schmidt Independence Criterion (HSIC) Lasso for model compression.
• A clustering-based pseudo-label generator to enhance learning from previously segmented classes.
• A customized soft labels module to maintain fine-grained knowledge of old classes.

Main Results:

• The proposed LISS model demonstrates superior performance compared to state-of-the-art methods on benchmark datasets.
• The LISS model achieves high effectiveness in continual semantic segmentation while significantly reducing computational and memory requirements.
• The pruning strategy and pseudo-label generation effectively mitigate catastrophic forgetting.

Conclusions:

• The LISS model offers an efficient and effective solution for class incremental semantic segmentation on edge devices.
• The developed methods for knowledge preservation and pseudo-labeling are crucial for successful continual learning in resource-limited scenarios.