MFNet: A Novel GNN-Based Multi-Level Feature Network With Superpixel Priors.

Area of Science:

• Computer Vision
• Machine Learning
• Graph Neural Networks

Background:

• Superpixel segmentation aggregates pixels by similarity, offering object boundary information.
• Existing methods struggle to effectively represent and leverage superpixel-level features for structural-aware learning.

Purpose of the Study:

• To propose a novel Multi-level Feature Network (MFNet) that effectively utilizes superpixel segmentation priors.
• To develop a robust superpixel representation method using graph neural networks.
• To demonstrate the versatility of MFNet across image-level and pixel-level prediction tasks.

Main Methods:

• Constructed graphs within and among superpixels to build the Multi-level Feature Network (MFNet).
• Employed a hierarchical approach to learn features from pixel-level to superpixel-level and then to image-level.
• Developed a superpixel representation method using graph neural networks for enhanced feature extraction.
• Designed specialized prediction modules (attention linear classifier, FC-based superpixel, Decoder-based pixel) for different tasks.

Main Results:

• MFNet achieved competitive results on various datasets compared to existing methods.
• Visualizations demonstrated more refined object boundaries and saliency map outlines.
• The network showed improved attention to details in predictions.

Conclusions:

• The proposed MFNet effectively leverages superpixel priors for structural-aware feature learning.
• MFNet offers a versatile framework applicable to diverse computer vision tasks.
• The novel superpixel representation method enhances feature extraction capabilities.