MFEAFN: Multi-scale feature enhanced adaptive fusion network for image semantic segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Semantic segmentation requires effective fusion of low-level spatial details and high-level semantic information.
• Existing methods face challenges in fully acquiring and integrating these diverse feature types.

Purpose of the Study:

• To propose a novel multiscale feature-enhanced adaptive fusion network (MFEAFN) for improved semantic segmentation performance.
• To enhance the extraction of high-level semantic information and the adaptive fusion of multi-scale features.

Main Methods:

• Designed a Double Spatial Pyramid Module (DSPM) for richer high-level semantic feature extraction.
• Developed a Focusing Selective Fusion Module (FSFM) utilizing spatial attention and 2D DCT for adaptive feature map fusion.
• Conducted comparative and ablation studies to validate the effectiveness of the proposed FSFM.

Main Results:

• The MFEAFN achieved 82.64% mIoU on PASCAL VOC2012 and 78.46% mIoU on Cityscapes.
• Demonstrated superior segmentation performance compared to existing state-of-the-art methods.
• Ablation studies confirmed the significant contribution of the FSFM to overall performance.

Conclusions:

• The proposed MFEAFN effectively fuses multi-scale features by enhancing spatial detail and semantic information.
• The DSPM and FSFM modules are crucial for achieving high performance in semantic segmentation.
• The MFEAFN represents a significant advancement in semantic segmentation technology.