飞机目标分类的深度卷积神经网络设计优化 
下载： 1184次
Design and Optimization of Deep Convolutional Neural Network for Aircraft Target Classification
河北工业大学电子信息工程学院, 天津 300401
图 & 表
图 1. 使用的6种类型的飞机目标。(a) Boeing;(b) Cessna172;(c) F/A18;(d) AH-64;(e) C-130;(f) MQ-9
Fig. 1. Six types of aircraft targets are used. (a) Boeing; (b) Cessna172; (c) F/A18; (d) AH-64; (e) C-130; (f) MQ-9
下载图片 查看原文
图 2. 飞机镜像操作效果图
Fig. 2. Effect of aircraft mirroring operation
下载图片 查看原文
图 3. 飞机旋转操作效果图
Fig. 3. Effect of aircraft rotation operation
下载图片 查看原文
图 4. 所设计的深度卷积神经网络结构图
Fig. 4. Structure of proposed deep convolutional neural network
下载图片 查看原文
图 5. 采用不同损失函数的DCNN性能随训练变化的曲线。(a)训练准确率;(b)验证准确率;(c)训练损失;(d)验证损失
Fig. 5. Curves of DCNN training performance by adopting different loss functions. (a) Train accuracy; (b) verification accuracy; (c) train loss; (d) verification loss
下载图片 查看原文
图 6. 训练集损失和验证集损失对比。(a)添加BN层;(b) dropout为0.5;(c) dropout为0.5,并添加BN层
Fig. 6. Comparison between train_loss and val_loss. (a) Adding BN layers; (b) dropout is 0.5; (c) dropout is 0.5, and BN layers are added
下载图片 查看原文
图 7. 飞机分类DCNN结构的归一化混淆矩阵
Fig. 7. Normalized confusion matrix of the proposed DCNN architecture for aircraft classification
下载图片 查看原文
表 1飞机型号参数列表
Table1. List of aircraft model parameters
| Aircraft type | Length /m | Height /m | Wing span range /m |
|---|
| Boeing | 46.61 | 12.92 | 44.42 | | Cessna172 | 8.28 | 2.72 | 11.00 | | F/A18 | 17.10 | 4.70 | 11.43 | | AH-64 | 17.73 | 3.87 | 14.63 | | C-130 | 29.79 | 11.66 | 40.41 | | MQ-9 | 11.00 | 3.80 | 20.00 |
|
查看原文
表 2不同卷积层数网络的分类性能和损失性能
Table2. Classification and loss performances of networks with different number of convolutional layers
| Number ofconvolutional layers | Classification accuracy | Loss |
|---|
| No. 1 | No. 2 | No. 3 | No. 1 | No. 2 | No. 3 |
|---|
| Four | 0.889 | 0.891 | 0.910 | 0.49 | 0.80 | 0.55 | | Five | 0.893 | 0.895 | 0.915 | 0.51 | 0.66 | 0.58 | | Six | 0.877 | 0.869 | 0.884 | 0.83 | 0.84 | 0.82 | | Seven | 0.858 | 0.859 | 0.870 | 1.38 | 1.84 | 1.30 |
|
查看原文
表 3不同池化方式的分类性能和损失性能
Table3. Classification and loss performances for different pooling methods
| Method of pooling | Classification accuracy | Loss |
|---|
| Max-pooling | 0.907 | 0.65 | | Average-pooling | 0.843 | 1.25 |
|
查看原文
表 4全连接层中神经元数量和隐藏层数的分类性能和损失性能
Table4. Classification and loss performances for the numbers of neurons and hidden layers in fully connected layer
| Numbers of hiddenlayers and neurons | Classificationaccuracy | Loss |
|---|
| Two (1024+1024) | 0.965 | 0.15 | | Two (1024+512) | 0.941 | 0.24 | | Three (1024+1024+1024) | 0.972 | 0.12 | | Three (1024+1024+512) | 0.978 | 0.15 |
|
查看原文
表 5采用不同优化器的分类性能
Table5. Classification performances of different optimizers
| Optimizer | Classification accuracy |
|---|
| SGD | 0.978 | | Adadelta | 0.594 | | RMSprop | 0.349 | | Adam | 0.173 |
|
查看原文
表 6三种减少过拟合方法的分类性能
Table6. Classification performances of three methods to reduce overfitting
| Method | Classification accuracy |
|---|
| BN layer | 0.936 | | Dropout is 0.5 | 0.912 | | BN layer,and dropout is 0.5 | 0.991 |
|
查看原文
表 7不同方法的识别效果对比
Table7. Comparison of different methods
| Method | Classification accuracy |
|---|
| AlexNet | 0.955 | | Proposed DCNN | 0.991 |
|
查看原文
马俊成, 赵红东, 杨东旭, 康晴. 飞机目标分类的深度卷积神经网络设计优化[J]. 激光与光电子学进展, 2019, 56(23): 231006. Juncheng Ma, Hongdong Zhao, Dongxu Yang, Qing Kang. Design and Optimization of Deep Convolutional Neural Network for Aircraft Target Classification[J]. Laser & Optoelectronics Progress, 2019, 56(23): 231006.
PDF全文
