基于改进CNN和GRU的遥感图像描述方法
Remote Sensing Image Description Method Based on Improved CNN and GRU
李铁柱 1张波 2沈夏炯 3魏金占 4党兰学 1郑逢斌3
作者信息
- 1. 河南大学河南省大数据分析与处理重点实验室,河南开封 475004;河南大学计算机与信息工程学院,河南开封 475004
- 2. 河南大学计算机与信息工程学院,河南开封 475004
- 3. 河南开封科技传媒学院信息工程学院,河南开封 475000
- 4. 桂林航天工业学院 电子信息与自动化学院,广西桂林 541004
- 折叠
摘要
为遥感图像生成可读的文本描述对于实现遥感图像的自动解译、高效挖掘遥感影像信息有着重要意义.传统卷积神经网络(Convolutional Neural Network,CNN)的全连接层只能接受固定输入,对尺度多变、背景复杂的遥感影像,会导致精度丢失.为此提出了一个改进的CNN和门控循环单元(Gated Recurrent Unit,GRU)相结合的模型,在CNN中加入了特征加权融合的空洞金字塔池化(Atrous Spatial Pyramid Pooling,ASPP)模块,可以在不丢失分辨率的情况下提高模型的精度.为了评估性能,在RSICD和UCM_Captions两个公开数据集进行实验.结果表明,该模型在遥感图像描述生成任务上表现优异,生成的自然语言描述更加准确,可以有效地提高遥感图像分析的自动化程度和语义准确性.
Abstract
Generating readable text descriptions for remote sensing images is of great significance for realizing automatic interpretation of remote sensing images and efficient mining of remote sensing image information.The fully connected layer of the traditional convolutional neural network(CNN)can only accept fixed inputs,remote sensing images with variable scales and complex backgrounds can lead to unnecessary loss of accuracy.We proposed a model that combines an improved CNN with a gated recurrent unit(GRU).The atrous spatial pyramid pooling(ASPP)module of feature weighted fusion is added to the CNN,which can be used to improve the accuracy of the model without losing resolution.In order to evaluate the performance,experiments were conducted on two public data sets,RSICD and UCM_Captions.The results show that the model performs well on the task of generating remote sensing image descriptions,and the generated natural language descriptions are more accurate,which can effectively improve the automation and semantic accuracy of remote sensing image analysis.
关键词
遥感图像/图像描述/卷积神经网络/门控循环单元/空洞金字塔池化Key words
remote sensing image/image description/convolutional neural network/LSTM/ASPP引用本文复制引用
基金项目
河南省重点研发计划科技攻关计划(232102210013)
出版年
2024
NETL
NSTL
万方数据