摘要
深度学习算法应用于SAR图像分类领域时存在模型训练时间较长且精度不够高等问题。对此,提出一种基于混合注意力机制的卷积神经网络模型,该模型基本模块分为主干分支和软分支。主干分支由残差收缩网络和改良之后的通道注意力机制组成,负责提取主要特征;软分支将下采样和上采样相结合,负责提取混合注意力权重,增强从输入到输出的映射能力。该模型在MSTAR数据集上取得了99.6%的识别率,且训练时间较短。噪声分析显示:该模型对椒盐噪声具有较强的鲁棒性。
When the deep learning algorithm is applied to the field of SAR image classification,there are some problems such as long model training time and low accuracy.In order to solve the problems,a convolution neural network model based on hybrid attention mechanism is proposed.The basic module of the model is divided into trunk branch and soft branch.The trunk branch is composed of residual shrinkage network and the improved channel attention mechanism,which is responsible for extracting the main features.The soft branch combines down sampling with up sampling to extract the hybrid attention weight and enhance the mapping ability from input to output.The recognition rate of the model is 99.6%on MSTAR data set,and the training time is short.Noise analysis shows that the model is robust to salt and pepper noise.
作者
史宝岱
张秦
李宇环
李瑶
SHI Baodai;ZHANG Qin;LI Yuhuan;LI Yao(Graduate College,Air Force Engineering University,Xi’an 710000,China)
出处
《电光与控制》
CSCD
北大核心
2023年第4期45-49,共5页
Electronics Optics & Control
基金
国家自然科学基金面上项目(61971438)。
关键词
SAR图像
深度学习
卷积神经网络
残差收缩网络
注意力机制
SAR image
deep learning
convolutional neural network
residual shrinkage network
attention mechanism