基于组合优化的遥感图文检索轻量化
A Lightweight Remote-sensing Image and Text Retrieval Approach Based on Combinatorial Optimization
赵良瑾 1卢宛萱 1于泓峰 1孙显1
作者信息
- 1. 中国科学院空天信息创新研究院,北京 100190;中国科学院网络信息体系重点实验室,北京 100190
- 折叠
摘要
针对现有基于图网络的遥感图文检索模型存在的海量参数、模型时效性低、存储空间需求大等问题,提出一种基于组合优化的遥感图文检索轻量化方法.从模型架构角度,设计基于跨阶段融合的轻量化卷积模块精简图文检索模型的参数;从数值量化角度,设计图网络混合精度训练与量化推理策略提升模型推理速度.在多个遥感检索数据集上的实验结果表明,该方法在检索精度基本不下降的条件下,总参数量、浮点运算量相比于典型方法降低60％以上.
Abstract
The existing graph network-based remote sensing image and text retrieval model has such problems as enormous parameters,low model timeliness,and large storage space requirements,etc.A lightweight remote sensing image and text retrieval approach based on combinatorial optimization is proposed.For model architecture,the parameters of simplified image and text retrieval model of lightweight convolutional module are designed based on a cross-stage fusion.For numerical quantification,the graph network mixed precision training and quantitative inference strategies are designed to increase the inference speed of the model.The experimental results on several remote sensing retrieval datasets show that,under the condition that accuracy is unaffected,the proposed method can reduce the total parameter quantity,floating point operations by over 60％compared with the typical method.
关键词
遥感图像/图文检索/图神经网络/轻量化模型Key words
remote sensing images/image and text retrieval/graph neural network/lightweight models引用本文复制引用
出版年
2024
NETL
NSTL
万方数据