基于多尺度注意力CNN的SAR遥感目标识别

Target Recognition Based on Multi-Scale Attention Convolutional Neural Network in SAR Remote Sensing Image

目标识别是合成孔径雷达(Synthetic Aperture Radar,SAR)图像解译的重要步骤。鉴于卷积神经网络(Convolutional Neural Network,CNN)在自然图像分类领域表现优越,基于CNN的SAR图像目标识别方法成为了当前的研究热点。SAR图像目标的散射特征往往存在于多个尺度当中,且存在固有的噪声斑,含有冗余信息,因此,SAR图像目标智能识别成为了一项挑战。针对以上问题,本文提出一种多尺度注意力卷积神经网络,结合多尺度特征提取和注意力机制,设计了基于注意力的多尺度残差特征提取模块,实现了高精度的SAR遥感图像目标识别。该方法在MSTAR数据集10类目标识别任务中的总体准确率达到了99.84%,明显优于其他算法。在测试集加入4种型号变体后,10类目标识别任务中的总体准确率达到了99.28%,验证了该方法在复杂情况下的有效性。

Target recognition is an important step in synthetic aperture radar(SAR)image interpretation.Due to the superior performance of convolutional neural network(CNN)in the field of natural image classification,target recognition methods in SAR image based on CNN have become a hotspot of current research.Scattering features of SAR image targets always exist in multiple scales,and there are inherent noise spots,which contain redundant information.Therefore,intelligent target recognition on SAR image has become a challenge.To solve the problems above,a model called multi-scale attention convolutional neural network is proposed.Combining multi-scale feature extraction and attention mechanism,a multi-scale feature extraction module based on attention is designed,which achieves high precision in target recognition of SAR remote sensing imagery.The overall accuracy of the proposed method can reach 99.84%in 10 types of target recognition tasks of MSTAR dataset,which is obviously better than other algorithms.After adding four subclasses of variants in the test set,the overall accuracy of 10 types of target recognition tasks can reach 99.28%,which verified the effectiveness of the algorithm in complex situations.