| 摘要: |
| 针对深度卷积神经网络训练需要大数量样本，采用迁移学习的方法辅助网络训练，解决了SAR图像样本不足的问题。通过控制对比实验，对每个卷积块权重进行迁移与分析，使用微调与冻结相结合的训练方式有效提高网络的泛化性与稳定性；然后根据目标检测任务的时效性对网络模型进行改进，提高了网络检测速度的同时减少了网络参数；最后结合复杂场景杂波切片对网络进行训练，降低了背景杂波的虚警目标数量，复杂多目标场景的检测结果表明所提出方法具有较好的检测性能。 |
| 关键词: 迁移学习 深度卷积神经网络 SAR目标检测 训练时间 |
| DOI：10.3969/j.issn.1672-2337.2018.05.011 |
| 分类号:TN958；TP183 |
| 基金项目: |
|
| Target Detection in SAR Images Based on Transfer Learning |
|
ZHANG Ye，ZHU Weigang
|
|
1.Graduate School，Space Engineering University，Beijing 101416，China;2.Department of Optical and Electronic Equipment，Space Engineering University，Beijing 101416，China
|
| Abstract: |
| To solve the problem that the training of deep convolution neural network requires a large number of samples，the transfer learning method is used to assist the small SAR image dataset for network training.By contrast experiments，the individual convolution weights are transferred and analyzed.The combination of the fine-tuned weights and the frozen weights is used to improve the generalization and stability of the network.Then，the network model is improved according to the target detection task.The network detection speed is increased and the network parameters are decreased.Finally，the complicated scene clutter slices are used to train the network.The number of false alarm targets under background clutter is reduced.The detection results of complex multi-target scenes show that the proposed method has better detection performance. |
| Key words: transfer learning deep convolution neural network SAR target detection training time |
