Congratulations to Xinfu Liu and Guangchen Shi on Their Paper Acceptance by IEEE IoTJ

Edge Computing Driven Active-Reference Fusion for Few-Shot Semantic Segmentation IEEE Internet of Things Journal

With the development of Few-shot Semantic Segmentation, it is now possible to predict various unseen categories using limited annotated data. However, existing methods based on large language models face high computational costs, making them insufficient and unreliable in specific scenarios. With the advancement of edge computing technology, most computational tasks can be allocated to edge servers, thereby reducing the computational burden on the entire system. To achieve efficiency, security, and flexibility, we propose an Edge Computing Active-Reference (ECAR) framework for few-shot segmentation. This framework includes a Mask Prediction Module (MPM) and an Iterative Fusion & Refinement Module (IFRM). Specifically, we propose an interactive segmentation strategy in MPM. This strategy not only accurately locates co-occurring symbiotic objects in support and query images but also relaxes the strict requirements for pixel-level annotations, allowing for weak boundary annotations. Based on the initial results computed by MPM, IFRM enhances feature channel information related to support images through a few-shot channel attention mechanism while iteratively refining segmentation masks to obtain more compact boundaries. In K-shot segmentation tasks, we further propose a Category-Modulation Module (CMM) to fuse features extracted from multiple annotated frames, thereby filtering out useless information and enhancing effective information. Experimental results show that the ECAR framework significantly improves the performance of few-shot semantic segmentation on edge devices, achieving mean Intersection over Union (mIoU) of 64.8% and 68.9% under 1-shot and 5-shot settings, respectively.