Dual-Branch Superpixel and Class-Center Attention Network for Efficient Semantic Segmentation.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Deep learning has advanced image semantic segmentation, but practical applications face challenges like real-time processing and accuracy.
• Existing methods often struggle with coarse edge segmentation, limited contextual understanding, and high computational costs.

Purpose of the Study:

• To develop an attention-guided dual-branch network for improved image semantic segmentation.
• To address limitations in edge accuracy, contextual understanding, and computational efficiency of current segmentation algorithms.

Main Methods:

• A superpixel sampling weighting module enhances boundary sensitivity and preserves local features by modeling pixel dependencies.
• A class-center attention module extracts class-specific features, reducing computational overhead and improving global feature representation.
• Learnable parameters adaptively fuse features from dual branches for focused information processing.

Main Results:

• The proposed method demonstrated superior segmentation performance on PASCAL VOC 2012, Cityscapes, and ADE20K datasets.
• Achieved better results compared to mainstream models like FCN, DeepLabV3+, and DANet using mIoU and PA metrics.
• The algorithm effectively balances segmentation accuracy with model efficiency.

Conclusions:

• The novel dual-branch network significantly enhances image semantic segmentation accuracy and efficiency.
• The attention-guided approach effectively tackles limitations of existing segmentation methods.
• This work provides a promising solution for real-time, high-accuracy image segmentation applications.