计算机科学
人工智能
图像分割
分割
模式识别(心理学)
计算机视觉
图像(数学)
尺度空间分割
特征提取
基于知识的系统
图像处理
基于分割的对象分类
图像纹理
数据挖掘
图像检索
图像匹配
作者
Qi Li,Guang Feng,Meng Li,Pengbo Si,Sijie Niu,Fei Richard Yu
标识
DOI:10.1109/tmm.2026.3743767
摘要
Knowledge distillation, as a model lightweighting technique, has been widely applied in various fields, such as image classification and image segmentation. However, its application to referring image segmentation remains unexplored. In this paper, we propose a cross-sample multi-modal knowledge distillation framework to achieve efficient knowledge transfer from teacher to student models, enabling high-performance yet lightweight referring image segmentation. First, we establish a simple yet effective vision-language deeply intertwined dual-stream encoding network as the foundational framework. Building upon this, we further introduce a cross-modal relational distillation loss to transfer the teacher model's visual-linguistic correlations to the student model. Finally, at the batch level, we explore cross-sample multi-scale relational distillation, guiding the student model to mimic the feature space distribution of the teacher model. Experimental results demonstrate that our proposed distillation method significantly enhances the student model's performance, even enabling some student models to surpass their teachers.
科研通智能强力驱动
Strongly Powered by AbleSci AI