一种改进的Capsule及其在SAR图像目标识别中的应用
An improved Capsule and its application in target recognition of SAR images
张盼盼 1,2,3,4,5,*罗海波 1,2,4,5鞠默然 1,2,3,4,5惠斌 1,2,4,5常铮 1,2,4,5
1 中国科学院沈阳自动化研究所,辽宁 沈阳 110016
2 中国科学院机器人与智能制造创新研究院,辽宁 沈阳 110169
3 中国科学院大学,北京 100049
4 中国科学院光电信息处理重点实验室,辽宁 沈阳 110016
5 辽宁省图像理解与视觉计算重点实验室,辽宁 沈阳 110016
图 & 表
图 1. Capsule单元结构
Fig. 1. Structure of Capsule unit
下载图片 查看原文
图 2. Capsule网络的重构层
Fig. 2. Layers of reconstruction of Capsule network
下载图片 查看原文
图 3. 原Capsule网络结构与改进的Capsule 网络结构
Fig. 3. Structure of original Capsule network and improved Capsule network
下载图片 查看原文
图 4. (a)和(b)分别为BMP2、BTR70、T72、BTR60和2S1的光学图像和相对应的SAR图像;(c)和(d)分别为BRDM2、D7、T62、ZIL131和ZSU23/4的光学图像和相对应得SAR图像
Fig. 4. Optical images and their corresponding MSTAR SAR images for (a) and (b) BMP2, BTR70, T72, BTR60, and 2S1; (c) and (d) BRDM2, D7, T62, ZIL131, and ZSU23/4
下载图片 查看原文
图 5. 改进的Capsule网络的重构结果。(a)原始图像,(b)目标图像,(c)重构图像
Fig. 5. Reconstruction result of improved Capsule. (a) Original image,(b) Target image and (c) Reconstruction image
下载图片 查看原文
图 6. 改进的Capsule网络训练中重构错误曲线和训练损失曲线。(a)重构错误曲线,(b)训练损失曲线
Fig. 6. Reconstruction error curve and training loss curve of improved Capsule network. (a) Reconstruction error curve and (b) training loss curve
下载图片 查看原文
表 1改进的Capsule网络与Capsule网络性能对比
Table1. Performance comparison of improved Capsule network and Capsule network
| Model size (parameters) | Training_time/epoch | BFLOPs | Capsule | 33.73 M | 2 min 8 s | 33.519 | Improved Capsule | 21.65 M | 1 min 3 s | 1.078 |
|
查看原文
表 2用于训练和测试实验的原始SAR数据集
Table2. Raw SAR dataset for training and testing in experiment
Class | BMP2sn-9563 | BTR70 | T72sn-132 | BTR60 | 2S1 | BRDM2 | D7 | T62 | ZIL131 | ZSU23/4 | Train samples(
)
| 117 | 117 | 116 | 128 | 150 | 149 | 150 | 150 | 150 | 150 | Test samples(
)
| 195 | 196 | 196 | 195 | 274 | 274 | 274 | 273 | 274 | 274 |
|
查看原文
表 3原Capsule网络10类目标识别结果的混淆矩阵(识别率:98.48%)
Table3. Confusion matrix of 10-class target recognition results of Capsule network(recognition rate: 98.48%)
Class | BMP2sn-9563 | BTR70 | T72sn-132 | BTR60 | 2S1 | BRDM2 | D7 | T62 | ZIL131 | ZSU23/4 | BMP2sn-9563 | 96.92 | 0.51 | 2.57 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | BTR70 | 0 | 100.00 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | T72sn-132 | 0 | 0 | 100.00 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | BTR60 | 0 | 0 | 0 | 98.46 | 0 | 0.51 | 0 | 0 | 0 | 1.03 | 2S1 | 0 | 0 | 0 | 2.92 | 94.16 | 1.46 | 0 | 0.73 | 0.73 | 0 | BRDM2 | 0 | 0.365 | 0 | 0.73 | 0 | 97.45 | 0 | 0 | 1.09 | 0.365 | D7 | 0 | 0.73 | 0 | 0 | 0 | 0 | 99.27 | 0 | 0 | 0 | T62 | 0 | 0 | 0 | 0 | 0.73 | 0 | 0 | 98.90 | 0 | 0.37 | ZIL131 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 100.00 | 0 | ZSU23/4 | 0 | 0 | 0 | 0 | 0 | 0 | 0.36 | 0 | 0 | 99.64 |
|
查看原文
表 4改进的Capsule网络10类目标识别结果的混淆矩阵(识别率:98.85%)
Table4. Confusion matrix of 10-class recognition results of improved Capsule network (recognition rate: 98.85%)
Class | BMP2sn-9563 | BTR70 | T72sn-132 | BTR60 | 2S1 | BRDM2 | D7 | T62 | ZIL131 | ZSU23/4 | BMP2sn-9563 | 96.41 | 0 | 3.59 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | BTR70 | 0 | 100.00 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | T72sn-132 | 0 | 0 | 100.00 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | BTR60 | 0 | 0 | 0 | 98.97 | 0 | 0 | 0 | 0 | 0 | 1.03 | 2S1 | 0 | 0 | 0 | 2.555 | 96.35 | 1.095 | 0 | 0 | 0 | 0 | BRDM2 | 0 | 0 | 0 | 0.73 | 0 | 98.54 | 0.365 | 0 | 0.365 | 0 | D7 | 0.365 | 0 | 0 | 0 | 0 | 0.365 | 99.27 | 0 | 0 | 0 | T62 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 99.63 | 0 | 0.37 | ZIL131 | 0 | 0 | 0 | 0 | 0 | 0 | 0.36 | 0 | 99.64 | 0 | ZSU23/4 | 0 | 0 | 0 | 0 | 0 | 0 | 0.36 | 0 | 0 | 99.64 |
|
查看原文
表 5不同方法的识别效果
Table5. Recognition performance of different methods
Methods | SOC | Rates | Training images | SVM[7] | 90.10% | 3 670 | AdaBoost[7] | 92.70% | 3 670 | DCNN[9] | 92.30% | 3 671 | DCNN[8] | 94.56% | 2 747 | IGT[7] | 95.00% | 3 670 | CGM[10] | 97.18% | 3 670 | 2-VDCNN[11] | 97.81% | 1 377 | CapsNet | 98.48% | 1 377 | Improved CapsNet | 98.85% | 1 377 |
|
查看原文
张盼盼, 罗海波, 鞠默然, 惠斌, 常铮. 一种改进的Capsule及其在SAR图像目标识别中的应用[J]. 红外与激光工程, 2020, 49(5): 20201010. Zhang Panpan, Luo Haibo, Ju Moran, Hui Bin, Chang Zheng. An improved Capsule and its application in target recognition of SAR images[J]. Infrared and Laser Engineering, 2020, 49(5): 20201010.