SAF-SD: Self-Distillation Object Segmentation Method Based on Sequential Three-Way Mask and Attention Fusion

Area of Science:

• Computer Vision
• Deep Learning
• Explainable AI

Background:

• Transformer models excel in computer vision but lack interpretability.
• Existing methods for Transformer interpretability produce coarse, noisy explanations due to mask limitations.

Purpose of the Study:

• To propose a self-distillation object segmentation method (SAF-SD) for improved Transformer model interpretability.
• To enhance salient and camouflaged binary object segmentation tasks.

Main Methods:

• Sequential three-way mask (S3WM) module for accurate foreground-background segmentation.
• Attention fusion (AF) module aggregating cross-layer attention for detail refinement and noise suppression.

Main Results:

• SAF-SD significantly improves object segmentation accuracy and explanation quality.
• The method effectively refines details and suppresses noise in explanation maps.

Conclusions:

• SAF-SD offers a more interpretable and reliable approach for Transformer-based computer vision.
• The proposed method addresses limitations of existing interpretation techniques, particularly for segmentation tasks.