基于多任务学习及Faster R-CNN的SAR目标图像识别分类

针对合成孔径雷达(SAR)目标图像识别分类中分类特征利用率低、精度差及图像特征提取时运算复杂、效率差的问题,利用非下采样剪切波变换(NSST)方向敏感性和平移不变性提取SAR目标图像的光谱纹理特征,构建基于Faster R-CNN(region-based convolutional neural network)的可同时完成目标图像识别、鉴别及分类的多任务网络模型。实验结果表明,该方法在有限的SAR图像数据支持下仍有较好的识别率,且算法优于传统的神经网络(NN)、支持向量机(SVM)及基于稀疏表示(Sc SPM)等分类方法。在MSTAR公开数据库上,平均识别率达到98.13%。

SAR target image recognition and classification based on Multitask learning and Faster Region-based Convolution Neural Network

For the questions of low utilization and precision in SAR image target recognition and classification ,the complicated operations and poor effect in the image feature extraction ,using the characteristics of the direction of sensitivity and the translation invariance of NSST for extracting the spectral texture features of SAR target image, proposing the method of target image recognition and classification based on the Multitask learning and Faster R-CNN. Building a network model of the multitask learning based on Faster R-CNN ,which can complete the recognition, identification and classification of SAR images at the same time .The results of the experiment show that this method still has a better recognition rate in the limited SAR images .And the algorithm is better than the traditional algorithm such as the NN ,SVM and Sc-SPM classification method. The average recognition rate can reach 98.13% in the publicly database of the MSTAR.