Article
Efficient Segmentation Using Attention-Fusion Modules With Dense Predictions
Serdar Erişen
IEEE Access, 2025, Volume 13, pp. 107552–107565
DOI: 10.1109/ACCESS.2025.3581986
This article extends the SERNet-Former research direction toward efficient semantic segmentation with attention-fusion modules and dense prediction mechanisms. It positions attention-fusion as a practical way to improve feature interaction, semantic propagation, and segmentation efficiency in CNN- and transformer-oriented dense prediction pipelines.
Research focus
Fusion modules for strengthening feature interaction and contextual reasoning.
Segmentation-oriented prediction paths designed for pixel-level outputs.
Architectural design for strong semantic segmentation with controlled complexity.
Module-level design direction compatible with modern convolutional and transformer-based vision models.
Keywords
Image segmentation; computer vision; semantic segmentation; attention-fusion mechanisms; convolutional neural networks; vision transformers; dense predictions.
BibTeX
@ARTICLE{11045846,
author={Erişen, Serdar},
journal={IEEE Access},
title={Efficient Segmentation Using Attention-Fusion Modules With Dense Predictions},
year={2025},
volume={13},
number={},
pages={107552-107565},
keywords={Semantics;Feature extraction;Transformers;Attention mechanisms;Immune system;Computer vision;Semantic segmentation;Logic gates;Convolutional neural networks;Head;Image segmentation;computer vision;attention-fusion mechanisms;convolutional neural networks;vision transformers;dense predictions},
doi={10.1109/ACCESS.2025.3581986}
}Connection to later work
SERNet-Former v2 forms part of the research path leading from attention-fusion modules and dense prediction toward the later NeuroFuser and NeuroFixer directions, where attention-fusion functions are made more modular, resource-aware, and reusable.