AGLN

AGLN enhances encoder-decoder image segmentation networks by combining attention-guided global enhancement and local refinement to improve semantic segmentation accuracy.

Key Features:

• Global Enhancement Method: Aggregates global information from high-level feature maps and adaptively distributes it across decoder layers to augment global context during upsampling, addressing limited local receptive fields of interpolation and deconvolution.
• Local Refinement Module: Uses decoder features as semantic guidance to refine noisy low-level encoder features from skip connections before fusion.
• Context Fusion Block: Integrates the Global Enhancement Method and the Local Refinement Module to harmonize global context enhancement with local feature refinement during feature fusion.

Scientific Applications:

• Semantic segmentation benchmarks: Evaluated on PASCAL Context, ADE20K, and PASCAL VOC 2012 for semantic segmentation tasks.
• ResNet-101 backbone performance: Using a vanilla ResNet-101 backbone, reported a mean Intersection over Union (IoU) of 56.23% on PASCAL Context.
• Comparative performance: Demonstrated improved performance over traditional encoder-decoder models in reported experiments.

Methodology:

Aggregates global information from high-level feature maps and adaptively distributes it across decoder layers during upsampling; refines low-level encoder features using decoder-derived semantic guidance before fusion; fuses enhanced global and refined local features within a Context Fusion Block; operates within encoder-decoder architectures with skip connections and upsampling methods such as interpolation and deconvolution.