SAR图像目标识别的卷积神经网模型

Convolution neural network model for SAR image target recognition

目的合成孔径雷达图像目标识别可以有效提高合成孔径雷达数据的利用效率。针对合成孔径雷达图像目标识别滤波处理耗时长、识别精度不高的问题,本文提出一种卷积神经网络模型应用于合成孔径雷达图像目标识别。方法首先,针对合成孔径雷达图像特点设计特征提取部分的网络结构;其次,代价函数中引入L2范数提高模型的抗噪性能和泛化性;再次,全连接层使用Dropout减小网络的运算量并提高泛化性;最后研究了滤波对于网络模型的收敛速度和准确率的影响。结果实验使用美国运动和静止目标获取与识别数据库,10类目标识别的实验结果表明改进后的卷积神经网络整体识别率(包含变体)由93. 76%提升至98. 10%。通过设置4组对比实验说明网络结构的改进和优化的有效性。卷积神经网络噪声抑制实验验证了卷积神经网络的特征提取过程对于SAR图像相干斑噪声有抑制作用,可以省去耗时的滤波处理。结论本文提出的卷积神经网络模型提高了网络的准确率、泛化性,无需耗时的滤波处理,是一种合成孔径雷达图像目标识别的有效方法。

Objective Synthetic aperture radar （SAR） is an important means of earth observation considering its all-weath- er, day-and-night, and penetrating imaging capabilities. SAR has been extensively used in battlefield detection and intelli- gence acquisition. SAR is a kind of electromagnetic wave coherent imaging system. A SAR image not only has variability but also has a strong speckle noise, which leads to considerable difficulties in target recognition of a SAR image. A manual interpretation of numerous SAR image data is difficult given the diversity of SAR image acquisition methods. A SAR auto- matic target recognition can effectively improve the utilization efficiency of SAR image data. However, the current SAR im- age target recognition algorithm has two main problems. First, the characteristics of target recognition, such as edge, cor- ner, contour, texture, and other low-level features, are not representative. Second, in the traditional SAR image target rec- ognition method, an effective filtering algorithm is crucial, but the filtering process is time-consuming. A convolutional neu- ral network model is presented in this study to solve the problems of time-consuming filtering process and low recognition ac- curacy in the SAR target recognition. Method First, a network structure of the feature extraction part was specifically de- signed for the characteristics of SAR images, which are slightly different from optical images. We must design a reasonable network structure for the characteristics of SAR images. First, a SAR image that reflects a target radar echo intensity is a gray image because the feature information is less in a SAR image than in an optical image. Second, speckle noise inevita- bly exists in the SAR image. Third, the pixel size of the target is small because of the resolution limitation of the SAR im-age. Owing to the characteristics of SAR images, the convolutional neural network applied to SAR image target recognition must use a small convolution kernel and an appropriate convolution layer number. The feature extraction part of the pro- posed convolutional neural network model consists of four convolutional layers, four nonlinear layers, and two pooling lay- ers. Second, an L2 norm was introduced to the cost function to improve the anti-noise and generalization performances of the model. Theoretical deduction shows the means by which the L2 norm enhances the noise immunity and generalization performance of the model. Third, Dropout reduced the computational complexity of the network and improved the generali- zation performance of the model. Dropout is a regularization technique for the reduction of overfitting in neural networks by preventing complex co-adaptations in training data. Dropout is an efficient technique for conducting model averaging with neural networks. Finally, the influence of filtering on the convergence speed and accuracy of the network was investigated. In the traditional SAR image target recognition method, the effective filtering algorithm is crucial, but the filtering process is time-consuming. Result Experimental data were obtained from the United States Moving and Stationary Target Acquisition and Recognition database. Experimental results of 10 types of target recognition showed that the overall recognition rate （ in- cluding the variant） of the improved convolutional neural network increased from 93.76% to 98. 10%. The improved fea- ture extraction network structure extracts effective target features, thus improving the accuracy of the model. The accuracy of target variant recognition in SAR images had also been considerably improved. Notably, L2 regularization and Dropout enhanced the generalization performance of the model. Different sets of comparative experiments were set up to illustrate the effectiveness of improving and optimizing the network structure. The accuracy rate decreased from 98. 10% to 97. 06% when the first layer uses a 9 ~ 9 convolution kernel instead of two cascaded 5 ~ 5 convolution kernels. The accuracy of net- work identification increases from 94. 91% to 96. 19% when using L2 regularization, thereby indicating that L2 regulariza- tion can effectively improve the accuracy of network identification. Dropout increases the fluctuation range of the recognition rate, thus increasing the recognition accuracy to the highest level. Noise suppression experiments on the convolutional neu- ral network were conducted to analyze the effects of three filtering methods, namely, Lee, bilateral, and Gamma MAP （ Maximum A Posteriori） , on the training process and results of the model. The experiments verified that the feature extrac- tion process of the convolutional neural network can suppress the speckle noise of the SAR image and can save time during the filtering process. The filtering process consumes additional time, does not improve the convergence speed of convolu- tional neural network training, and decreases the recognition accuracy because it may filter out effective target recognition features, such as target texture, thus resulting in a decrease in recognition accuracy. Conclusion The convolutional neural network model proposed in this study improves the accuracy and generalization of the network, does not require a time-con- suming filtering process, and is an effective method for target recognition of SAR images.