Edge-Computing-Driven Active-Reference Fusion for Few-Shot Semantic Segmentation

doi:10.1109/JIOT.2025.3587649

Edge-Computing-Driven Active-Reference Fusion for Few-Shot Semantic Segmentation

2025-10-28

0

PRE

AI

DOI：10.1109/JIOT.2025.3587649

Original

Original request for help

Save

Abstract

En 中文

With the development of few-shot semantic segmentation, various unseen classes are predicted via few labeled data. However, existing methods suffer from heavy computation costs, making them inadequate and unreliable in specific scenarios. As edge computing advances, computation can be allocated to edge servers, largely relieving the computational pressure. We thus, propose edge computing active-reference (ECAR) framework for few-shot segmentation, including mask prediction module (MPM) and iterative fusion and refinement module (IFRM). Specifically, mutual segmentation strategy is proposed in MPM, which not only accurately locates co-occurrence objects appearing in both support and query images, but also relaxes high constraints on pixel-level labeling, allowing for weak boundary labeling. Based on results computed by MPM, IFRM enhances feature channel information related to the supported image via few-shot channel attention scheme, iteratively refining segmentation masks to obtain a compact boundary. In k-shot segmentation, we propose category-modulation module to fuse features extracted from multiple annotated frames, thus forgetting useless information and enhancing contributive information. Experiments show edge computing driven active-reference (ECAR) boosts segmentation on edge devices, achieving 64.8% and 68.9% of m-IOU in 1-shot and 5-shot, respectively.

Keywords:

Edge computing

semantic segmentation

Edge-Computing-Driven Active-Reference Fusion for Few-Shot Semantic Segmentation

Abstract

Journal

Organization