目的 深度学习已经大量应用于合成孔径宽达（SAR）图像目标识别领域，但大多数工作是基于MSTAR数据集的标准操作条件展开研究。当将深度学习应用于同类含变体目标时，例如T72子类，由于目标间差异小，所以仍存在着较大的挑战。本文从极大限度地保留SAR图像输入特征出发，设计一种适用于SAR变体目标识别的深度卷积神经网络结构。方法 设计网络主要由多尺度空间特征提取模块和DenseNet中的稠密块、转移层构成。多尺度特征提取模块置于网络底层，通过使用尺寸分别为1×1、3×3、5×5、7×7、9×9的卷积核，提取丰富空间特征的同时保留输入图像信息。为使输入图像信息更加有效地向后传递，基于DenseNet中的稠密块和转移层进行后续网络层设计。在对训练样本进行样本扩充基础上，分析了输入图像分辨率及目标存在平移和不同噪声水平等情况对模型识别精度的影响，与用于SAR图像目标识别的深度模型识别精度在标准操作条件下进行了对比分析。结果 实验结果表明，对T72 8类变体目标进行分类，设计的模型能够取得95.48%的识别精度，在存在目标平移和不同噪声水平情况下，平均识别精度分别达到了94.61%和86.36%。对10类目标（包括不含变体和含变体情况）在进行数据增强的情况下进行模型训练与测试，分别达到了99.38%和98.81%的识别精度，略优于其他对比模型结构识别精度。结论 提出的模型可以充分利用输入图像以及各卷积层输出的特征，学习目标图像的细节差异，不仅适用于SAR图像变体目标的识别任务，同时在标准操作条件下的识别任务也取得了较高的识别结果。

Objective Deep learning has been widely used in the field of synthetic aperture radar (SAR) target recognition and most studies have been conducted for target recognition under the standard operating conditions (SOCs) of MSTAR datasets. Many challenges exist due to the small differences among the targets when applied to target recognition with variants, such as T72 subclasses. To preserve the input features of SAR images, a deep convolutional neural network (CNN) architecture for SAR target recognition with variants is designed in this study. Method The proposed network is composed of one multiscale feature extraction module and several dense blocks and transition layers proposed in DenseNet. The multiscale feature extraction module, which is placed at the bottom of the network, uses multiple convolution kernels with sizes of 1×1, 3×3, 5×5, 7×7, and 9×9 to extract rich spatial features. The convolution kernels with a size of 1×1 are adopted to preserve the detailed information from the input image, and convolution kernels with large sizes are used in multiscale feature extraction module to suppress the influence of speckle noise on extracted features because speckle noise is a main factor that affects recognition performance. To transfer the information from the input image effectively and utilize the feature learned from all layers, dense blocks and transition layers are adopted in designing the latter layers of the network. A full convolution layer is used behind three dense blocks and transition layers to transform the learned features to vectors, and a SoftMax layer is adopted to perform classification. Finally, training datasets are augmented by displacing and adding speckle noise to the original images, and the proposed model is implemented using TensorFlow and is trained by using these samples. The influences of input image resolution, target translation, and different noise levels on the recognition accuracy of the proposed network are determined after augmenting the training datasets, and performance comparisons with other deep learning models under SOCs. Result Experimental results demonstrate that the input image resolution has a considerable influence on the recognition accuracy for eight types of T72 targets, and the accuracy improves considerably with the increase of input resolution. However, the input resolution has minimal effect on the recognition accuracy for SOC due to the large differences among the targets in SOC. The image resolution as the input of the proposed model is set to 88×88×1 because the target and shadow information during data enhancement should be preserved. To verify the performance of the proposed multiscale feature extraction module, tests are performed using different multiscale feature extraction strategies, and the proposed model obtains a classification accuracy of approximately 95.48% in the classification of eight subclasses of T72 target with variants. Aside from the recognition of test samples under SOC, the classification accuracies of the proposed model are investigated in terms of target translation and different noise levels. The proposed model can achieve a recognition accuracy higher than 90%, especially when the target is displaced 16 pixels away from the center of the original image. The proposed model still exhibits a good performance when the noise intensity is set to 0.5 or 1 but causes a remarkable decline in recognition accuracy when the noise intensity is greater than 1. The average classification accuracy can reach 94.61% and 86.36% in the case of object translation and different noise levels. Recognition accuracies of 99.38% (SOC1-10), 99.50% (SOC1-14), and 98.81% (SOC2) are achieved by using augmented training datasets in training the models for 10-class target recognition under SOC (without variants and with variants). Our model achieves comparable recognition performance with other deep models. Conclusion Our model utilizes the input information and features of each convolutional layer and captures the detailed difference among the targets from the images. Our model not only can be applied to target recognition task with variants but also achieve satisfactory recognition results under SOC.