Ballistic target recognition based on cost-sensitively pruned convolutional neural network
-
摘要:
为降低弹道目标整体误识别代价，提出了基于代价敏感剪枝（CSP）一维卷积神经网络（1D-CNN）的弹道目标高分辨距离像识别方法。首先，基于彩票假设提出了同时以降低模型复杂度和误识别代价为目标的统一框架；然后，在此基础上，提出了基于人工蜂群算法的网络结构无梯度优化方法，以网络结构搜索的方式自动地寻找1D-CNN的代价敏感子网络，即代价敏感剪枝；最后，为了使代价敏感子网络在微调过程中仍以最小化误识别代价为目标，提出了一种代价敏感交叉熵（CSCE）损失函数对训练进行优化，使代价敏感子网络侧重对误识别代价较高的类别正确分类来进一步降低整体误识别代价。实验结果表明：结合CSP和CSCE损失函数的1D-CNN能在保持较高的识别正确率的前提下，相比传统的1D-CNN具有更低的整体误识别代价，且降低了50%以上的计算复杂度。
Abstract:Aimed at reducing the overall misrecognition cost of ballistic targets, A One-Dimensional Convolutional Neural Network (1D-CNN) based on Cost-Sensitively Pruning (CSP) is proposed for ballistic target high-resolution range profile recognition. Firstly, based on the lottery ticket hypothesis, a unified framework is proposed to reduce the model complexity and overall misidentification cost concurrently. On this basis, a gradient-free optimization method of network structure based on artificial bee colony algorithm is proposed, which can automatically find the cost-sensitive subnetwork of 1D-CNN, namely, cost-sensitively pruning. Finally, in order to make the cost-sensitive sub-network still be aimed at minimizing the cost of misrecognition during the fine-tuning process, a novel Cost-Sensitive Cross Entropy (CSCE) loss function is proposed to optimize the training, so that the cost-sensitive sub-network focuses more on correctly classifying the categories with higher misrecognition cost to further reduce the overall misrecognition cost. The experimental results show that the proposed 1D-CNN combined with the CSP and CSCE loss function has a lower overall misrecognition cost than traditional 1D-CNN under the premise of maintaining a higher recognition accuracy, and reduces the computational complexity by more than 50% as well.
-
图 4 代价矩阵为M1时4种模型在不同数据集上的训练曲线
Figure 4. Training curves of four models in different datasets when cost matrix is M1
图 5 代价矩阵为M1时4种模型在不同SNR数据集上的整体误识别代价
Figure 5. Total misrecognition cost of four models in different datasets with different SNR when cost matrix is M1
图 6 代价矩阵为M1时超参数λ对实验结果的影响
Figure 6. Effects of hyper-parameter λ on experimental results when cost matrix is M1
图 7 代价矩阵为M1时超参数α对实验结果的影响
Figure 7. Effects of hyper-parameter α on experimental results when cost matrix is M1
表 1 数据集样本数量
Table 1. Sample number of datasets
数据集 训练数据集各类样本数 测试数据集各类样本数 弹头 高仿诱饵 简单诱饵 母舱 球形诱饵 Im0 2 881 2 881 2 881 2 881 2 881 720 Im1 2 305 2 449 2 593 2 737 2 881 720 Im2 1 729 2 017 2 305 2 593 2 881 720 Im3 1 152 1 729 2 305 2 593 2 881 720 
下载: 导出CSV
表 2 四种方法的识别结果
Table 2. Recognition results of four methods
代价矩阵 数据集 测试数据集整体误识别代价 测试数据集整体识别正确率/% CNN1D(CE) CNN1D(CSP+CE) CNN1D(CSCE) CNN1D(CSP+CSCE) CNN1D(CE) CNN1D(CSP+CE) CNN1D(CSCE) CNN1D(CSP+CSCE) M1 Im0 924.00±26.00 896.00±42.00 762.00±48.00 648.00±68.00 95.29±0.26 95.50±0.12 94.47±0.44 95.38±0.53 Im1 830.00±52.00 1 016.00±28.00 762.00±14.00 742.00±18.00 95.50±0.35 95.50±0.70 94.74±0.12 95.47±0.09 Im2 1 012.00±78.00 1 190.00±112.00 852.00±28.00 771.00±15.00 94.33±1.20 95.58±0.32 94.12±0.26 95.03±0.29 Im3 1 527.00±17.00 1 574.00±110.00 1 012.00±62.00 857.00±128.00 92.98±0.82 93.60±0.56 93.51±0.23 93.92±0.82 M2 Im0 142.05±10.45 176.20±5.00 142.10±12.60 136.75±6.25 95.76±0.03 95.96±0.18 94.50±0.41 94.88±0.03 Im1 141.30±3.00 169.65±4.65 142.65±7.55 134.75±0.45 95.44±0.18 96.05±0.50 94.74±0.29 94.94±0.09 Im2 175.85±9.25 197.70±15.30 151.20±3.20 143.60±8.80 95.35±0.03 95.32±0.06 94.24±0.20 94.50±0.12 Im3 209.65±22.85 226.64±1.65 179.30±14.60 174.70±0.30 94.42±0.03 94.15±0.53 93.01±0.61 94.44±0.06 M3 Im0 845.45±74.05 890.10±21.30 838.75±22.95 811.35±13.05 95.44±0.18 95.99±0.03 94.91±0.18 95.50±0.24 Im1 803.70±50.40 968.15±13.45 939.15±42.55 769.65±32.15 95.88±0.44 96.26±0.18 94.30±0.32 95.85±0.12 Im2 910.70±16.90 1 145.80±31.00 979.00±40.60 922.20±31.70 95.44±0.06 95.56±0.12 94.27±0.29 94.92±0.35 Im3 1 342.00±71.00 1 346.75±29.85 1 155.10±122.30 1 122.25±4.75 94.12±0.20 94.39±0.18 93.36±0.67 93.42±0.15
下载: 导出CSV
表 3 三种指标下模型剪枝量百分比
Table 3. Pruned percentages of model under three metrics
代价矩阵 数据集 浮点运算量/% 参数总数/% 通道总数/% M1 Im0 75.90±2.88 83.10±7.20 58.00±12.07 Im1 65.93±11.19 80.99±5.56 60.13±4.73 Im2 65.99±13.20 78.21±3.00 54.80±5.53 Im3 65.58±9.18 70.43±6.63 48.43±7.77 M2 Im0 61.89±13.20 54.58±10.18 35.60±7.00 Im1 81.93±5.18 86.56±4.95 60.63±1.63 Im2 53.74±3.46 57.30±7.38 34.50±6.37 Im3 72.52±0.90 82.45±6.84 57.70±10.17 M3 Im0 67.36±3.13 77.97±2.69 49.57±9.50 Im1 67.21±5.24 83.97±5.20 61.87±4.07 Im2 63.61±3.86 74.90±12.96 52.67±15.80 Im3 62.81±3.58 69.81±10.60 45.73±13.73
下载: 导出CSV
-
[1] PERSICO A R, ILIOUDIS C V, CLEMENTE C, et al. Novel classification algorithm for ballistic target based on HRRP frame[J]. IEEE Transactions on Aerospace and Electronic Systems, 2019, 55(6): 3168-3189. doi: 10.1109/TAES.2019.2905281 [2] 赵振冲, 王晓丹. 引入拒识的最小风险弹道目标识别[J]. 西安交通大学学报, 2018, 52(4): 132-138. https://www.cnki.com.cn/Article/CJFDTOTAL-XAJT201804019.htmZHAO Z C, WANG X D. A minimum risk recognition method of ballistic targets with rejection options[J]. Journal of Xi'an Jiaotong University, 2018, 52(4): 132-138(in Chinese). https://www.cnki.com.cn/Article/CJFDTOTAL-XAJT201804019.htm [3] KRIZHEVSKY A, SUTSKEVER I, HINTON G E. ImageNet classification with deep convolutional neural networks[C]//Proceedings of the 25th International Conference on Neural Information Processing Systems, 2012, 1: 1097-1105. [4] CHEN W, WANG Y, SONG J, et al. Open set HRRP recognition based on convolutional neural network[J]. The Journal of Engineering, 2019, 19(21): 7701-7704. http://ieeexplore.ieee.org/document/8879042 [5] GUO C, HE Y, WANG H P, et al. Radar HRRP target recognition based on deep one-dimensional residual-inception network[J]. IEEE Access, 2019, 7(2): 9191-9204. http://www.onacademic.com/detail/journal_1000041624730599_3db9.html [6] WAN J, CHEN B, XU B, et al. Convolutional neural networks for radar HRRP target recognition and rejection[J]. EURASIP Journal on Advances in Signal Processing, 2019, 5(19): 1-27. doi: 10.1186/s13634-019-0603-y [7] WEN Y, SHI L C, YU X, et al. HRRP target recognition with deep transfer learning[J]. IEEE Access, 2020, 8(22): 57859-57867. http://ieeexplore.ieee.org/document/9040527 [8] XIANG Q, WANG X D, SONG Y F, et al. One-dimensional convolutional neural networks for high-resolution range profile recognition via adaptively feature recalibrating and automatically channel pruning[J]. International Journal of Intelligent Systems, 2021, 36(1): 332-361. doi: 10.1002/int.22302 [9] SANTURKAR S, TSIPRAS D, ILYAS A, et al. How does batch normalization help optimization [EB/OL]. (2018-05-29)[2020-08-14]. https://arxiv.org/abs/1805.11604. [10] IOFFE S, SZEGEDY C. Batch normalization: Accelerating deep network training by reducing internal covariate shift[EB/OL]. (2015-02-11)[2020-08-14]. https://arxiv.org/abs/1502.03167. [11] MISRA D. Mish: A self regularized non-monotonic neural activation function[EB/OL]. (2019-08-23)[2020-08-14]. https://arxiv.org/abs/1908.08681. [12] GOHIL V, NARAYANAN S D, JAIN A. One ticket to win them all: Generalizing lottery ticket initializations across datasets and optimizers[EB/OL]. (2019-06-06)[2020-08-14]. https://arxiv.org/abs/1906.02773. [13] FRANKLE J, CARBIN M. The lottery ticket hypothesis: Finding sparse, trainable neural networks[EB/OL]. (2018-03-09)[2020-08-14]. https://arxiv.org/abs/1803.03635. [14] LIU Z, SUN M, ZHOU T, et al. Rethinking the value of network pruning[J/OL]. (2018-10-11)[2020-08-14]. https://arxiv.org/abs/1810.05270. [15] 李秋洁, 赵亚琴, 顾洲. 代价敏感学习中的损失函数设计[J]. 控制理论与应用, 2015, 32(5): 689-694. https://www.cnki.com.cn/Article/CJFDTOTAL-KZLY201505015.htmLI Q J, ZHAO Y Q, GU Z. Design of loss function for cost-sensitive learning[J]. Control Theory & Applications, 2015, 32(5): 689-694(in Chinese). https://www.cnki.com.cn/Article/CJFDTOTAL-KZLY201505015.htm [16] 向前, 王晓丹, 李睿, 等. 基于DCNN的弹道中段目标HRRP图像识别[J]. 系统工程与电子技术, 2020, 42(11): 2426-2433. https://www.cnki.com.cn/Article/CJFDTOTAL-XTYD202011004.htmXIANG Q, WANG X D, LI R, et al. HRRP image recognition of midcourse ballistic targets based on DCNN[J]. Systems Engineering and Electronics, 2020, 42(11): 2426-2433(in Chinese). https://www.cnki.com.cn/Article/CJFDTOTAL-XTYD202011004.htm [17] DUBEY S R, CHAKRABORTY S, ROY S K, et al. diffGrad: An optimization method for convolutional neural networks[J]. IEEE Transactions on Neural Networks and Learning Systems, 2020, 55(2): 1-12. http://ieeexplore.ieee.org/document/8939562 -
下载:
点击查看大图
计量
- 文章访问数: 500
- HTML全文浏览量: 130
- PDF下载量: 68
- 被引次数: 0
