MGSeg: Multiple Granularity-Based Real-Time Semantic Segmentation Network.

Area of Science:

• Computer Vision
• Deep Learning

Background:

• Semantic segmentation performance is enhanced by global contextual information.
• Real-time processing remains a challenge for complex segmentation tasks.

Purpose of the Study:

• To propose an efficient multi-granularity based semantic segmentation network (MGSeg) for real-time applications.
• To model the latent relevance between multi-scale geometric details and high-level semantics for fine-grained segmentation.

Main Methods:

• Utilized a light-weight ResNet-18 backbone for hierarchical feature extraction.
• Introduced Hybrid Attention Feature Aggregation (HAFA) to filter noisy details and improve feature representation.
• Employed a Fine Granularity Refinement (FGR) module to model multi-level feature-category relationships for weighted fusion.
• Implemented light-weight strategies like channel compression and narrow neck structures for efficiency.

Main Results:

• Achieved state-of-the-art performance on Cityscapes (77.8%@50fps) and CamVid (72.7%@127fps) datasets.
• Demonstrated significant improvements in real-time semantic segmentation accuracy and speed.
• Validated the effectiveness of the proposed MGSeg network for practical, time-sensitive applications.

Conclusions:

• The MGSeg network effectively balances accuracy and efficiency for real-time semantic segmentation.
• The proposed HAFA and FGR modules contribute to improved feature learning and segmentation refinement.
• MGSeg shows strong potential for deployment in real-world autonomous systems requiring rapid scene understanding.