计算机科学
人工智能
自然语言处理
分解
情报检索
可视化
语义学(计算机科学)
模式识别(心理学)
矩阵分解
视觉文字
图像检索
语义压缩
视频检索
语义鸿沟
语义相似性
语义计算
作者
Donglin Zhang,Zhiwen Wang,Xiao‐Jun Wu
标识
DOI:10.1109/tmm.2026.3719812
摘要
In recent years, text-video retrieval (TVR) tasks have attracted widespread attention. Existing approaches usually map video clips and text descriptions into a joint semantic embedding space to facilitate cross-modal interaction and alignment. However, due to the inherent differences between video and text modalities, video contents are often redundant and complex, containing repetitive and irrelevant noise compared to text descriptions. Consequently, these distractions may adversely affect retrieval performance during semantic interaction and alignment. To this end, we propose a novel visual semantic decomposition and textual concept enhancement method for text-video retrieval tasks, named SDCE. Specifically, the proposed SDCE introduces a visual semantic decomposition module to segregate semantic contents in videos, reducing the impact of noisy and redundant information. Besides, the Gram regularization scheme is developed to ensure the independence of decomposed contents. To enhance the global textual features, the concept enhancement module is proposed by integrating core concepts from Top-k word tokens. Finally, we conduct extensive experiments on three prominent datasets (i.e., MSRVTT, MSVD, DiDeMo and ActivityNet), demonstrating the effectiveness of our proposed method.
科研通智能强力驱动
Strongly Powered by AbleSci AI