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。
Research on Application of Convolutional Neural Network in SAR Image Target Recognition
LIN Zhilong,WANG Changlong,HU Yongjiang,ZHANG Yan(Department of Unmanned Aerial Vehicle,Army Engineering University)
Objective SAR is one of the most important means of earth observation because of its all-weather, day-and-night and penetrating imaging capability. SAR has been widely used in battlefield detection and intelligence acquisition. SAR is a kind of electromagnetic wave coherent imaging system. Its image not only has variability, but also has strong speckle noise, which brings great difficulties to target recognition of SAR image. In addition, due to the diversity of SAR image acquisition methods, it is difficult to manually interpret a large number of SAR image data. SAR Automatic Target Recognition (ATR) can effectively improve the utilization efficiency of SAR image data. However, there are two main problems in the current SAR image target recognition algorithm: first, the characteristics of the target recognition are not representative, such as edge features, corner features, contour, texture and other features, which are low level features. Second, in the traditional SAR image target recognition method, the effective filtering algorithm is very important, but the filtering process is very time-consuming. In order to solve the problem of time-consuming filter and low recognition accuracy in SAR targets recognition, a convolutional neural network model was presented in this paper.Method First of all, the network structure of the feature extraction part was designed for the characteristics of synthetic aperture radar images. SAR images are quite different from optical images. We must design a reasonable network structure for the characteristics of SAR images. First, the SAR image that reflects the target radar echo intensity is a gray image. Compared with the optical image, the feature information is less; second, the speckle noise inevitably exists in the SAR image; and third, the pixel size of the target is smaller because of the resolution limitation of the SAR image. In view of the above SAR image characteristics, convolution neural network applied to SAR image target recognition should use smaller convolution kernel and appropriate convolution layer number. The feature extraction part of the proposed convolution neural network model consists of 4 convolutional layers, 4 nonlinear layers and 2 pooling layers. Secondly, the L2 norm was introduced into the cost function to improve the anti-noise performance and generalization of the model. The theoretical deduction shows how the L2 norm enhances the noise immunity and generalization performance of the model. Thirdly, Dropout reduced the computational complexity of the network and improved the generalization of the model. Dropout is a regularization technique for reducing overfitting in neural networks by preventing complex co-adaptations on training data. It is a very efficient way of performing model averaging with neural networks. Finally, the influence of filtering on the convergence speed and accuracy of the network was studied. In the traditional SAR image target recognition method, the effective filtering algorithm is very important, but the filtering process is very time-consuming.Result Experimental data were the United States Moving and Stationary Target Acquisition and Recognition database. The experimental results of 10 types of target recognition showed that the overall recognition rate (including variant) of the improved convolution neural network is raised from 93.76% to 98.10%. The improved feature extraction network structure extracts more effective target features, thus improving the accuracy of the model. The accuracy of target variants recognition in SAR images has also been greatly improved. It showed that L2 regularization and Dropout enhance the generalization performance of the model. a sets of comparative experiments were set up to illustrate the effectiveness of the improvement and optimization of network structure. When the first layer uses 9×9 convolution kernel instead of two cascaded 5×5 convolution kernel, the accuracy rate decreases from 98.10% to 97.06%, and the accuracy rate decreases. The accuracy of network identification is increased from 94.91% to 96.19% by using L2 regularization, which shows that L2 regularization can effectively improve the accuracy of network identification. Dropout increases the fluctuation range of recognition rate when increasing the highest recognition accuracy. The noise suppression experiments of convolution neural network were conducted to study the effects of three filtering methods, Lee filtering, bilateral filtering and Gamma MAP filtering, on the training process and results of the model. It verified that the feature extraction process of the convolution neural network can suppress the speckle noise of the SAR image, and can save the time-consuming filtering processing. Filtering processing takes longer time and does not improve the convergence speed of convolution neural network training, but the recognition accuracy decreases. This is because filtering process may filter out effective target recognition features such as target texture, resulting in a decrease in recognition accuracy.Conclusion The convolution neural network model proposed in this paper improves the accuracy and generalization of the network, and does not need time-consuming filtering processing. It is an effective method for target recognition of SAR images.