结合卷积神经网络多层特征和支持向量机的车辆识别 
下载： 1361次
Vehicle Recognition Based on Multi-Layer Features of Convolutional Neural Network and Support Vector Machine
西北师范大学物理与电子工程学院, 甘肃 兰州 730070
图 & 表
图 1. 基于MCP-SVM混合模型的车辆识别算法架构
Fig. 1. Structure of vehicle recognition method based on MCP-SVM hybrid model
下载图片 查看原文
图 2. 部分图片样本示例。(a)正样本;(b)负样本
Fig. 2. Several images of samples. (a) Positive samples; (b) negative samples
下载图片 查看原文
图 3. 正确率和训练损失曲线对比图
Fig. 3. Comparison of accuracies and training loss curves
下载图片 查看原文
表 1基于AlexNet的7种网络结构
Table1. Seven kinds of network structures based on AlexNet
| Networkname | Imageinput | Convolutionkernel | Networklayer | C1 | S1 | C2 | S2 | C3 | S3 | C4 | C5 | S5 | C6 |
|---|
| Ⅰ | 28×28 | 3 | 8 | 3×3 | 2×2 | 2×2 | 2×2 | 3×3 | - | 2×2 | 2×2 | 2×2 | - | | Ⅱ | 48×48 | 5 | 8 | 5×5 | 2×2 | 5×5 | 2×2 | 4×4 | - | 3×3 | 3×3 | 2×2 | - | | Ⅲ | 96×96 | 5 | 8 | 5×5 | 2×2 | 5×5 | 2×2 | 5×5 | 2×2 | 5×5 | 5×5 | 2×2 | - | | Ⅳ | 28×28 | 5 | 8 | 5×5 | 2×2 | 3×3 | 2×2 | 2×2 | - | 2×2 | 2×2 | 2×2 | - | | Ⅴ | 28×28 | 7 | 8 | 7×7 | 2×2 | 2×2 | 2×2 | 2×2 | - | 2×2 | 2×2 | 2×2 | - | | Ⅵ | 28×28 | 3 | 7 | 3×3 | 2×2 | 2×2 | 2×2 | 3×3 | 2×2 | 2×2 | - | - | - | | Ⅶ | 28×28 | 3 | 9 | 3×3 | 2×2 | 2×2 | - | 2×2 | 2×2 | 2×2 | 3×3 | - | 2×2 |
|
查看原文
表 27种网络的分类性能
Table2. Classification performance of seven kinds of networks
| Classification network | Ⅰ | Ⅱ | Ⅲ | Ⅳ | Ⅴ | Ⅵ | Ⅶ |
|---|
| Training time /h | 5.8 | 11.5 | 68 | 5.2 | 10 | 6.3 | 7.5 | | Accuracy rate /% | 97.82 | 97.62 | 94.00 | 96.92 | 96.72 | 97.87 | 97.76 |
|
查看原文
表 3改进的CNN模型结构
Table3. Structure of improved CNN model
| Layer | Layer input | Convolution kernal | Convolution output | Pooling | Pooled output |
|---|
| Size | | Num | Step | Size | Mode |
|---|
| L1(C1+S1) | 28×28×3 | 3×3 | 96 | 1 | 26×26×96 | 2×2 | Max | 13×13×96 | | L2(C2+S2) | 13×13×96 | 2×2 | 128 | 1 | 12×12×128 | 2×2 | Max | 6×6×128 | | L3(C3+S3) | 6×6×125 | 3×3 | 256 | 1 | 4×4×256 | 2×2 | Max | 2×2×256 | | L4(C4) | 2×2×256 | 2×2 | 256 | 1 | 1×1×256 | - | - | - | | L5(Fc1) | 1×1×256 | - | - | - | - | - | - | 1024 | | L6(Fc2) | 1024 | - | - | - | - | - | - | 1024 | | L7(Softmax) | 1024 | - | - | - | - | - | - | 2 |
|
查看原文
表 4CNN模型对比分析
Table4. Comparison and analysis of CNN models
| Method | Training time /h | Accuracy rate /% |
|---|
| Using AlexNet model | 51 | 96.92 | | Using improved model | 6.3 | 97.87 |
|
查看原文
表 53种方法分类性能对比
Table5. Comparison of classification performance of three methods
| No. | Method | Accuracyrate /% | Testingtime /s |
|---|
| 1 | Method in Ref. [20] | 98.32 | 88.36 | | 2 | MC-SVM | 98.72 | 247.51 | | 3 | MCP-SVM | 98.73 | 13.19 |
|
查看原文
表 6不同方法在车辆数据集的识别率对比
Table6. Comparison of recognition rates of different methods in vehicle datasets
| Method | Accuracyrate /% | Testingtime /s |
|---|
| Improved CNN model | 97.87 | 184 | | Method in Ref. [21] | 91.75 | 1596 | | Method in Ref. [4] | 92.33 | 1046 | | Method in Ref. [6] | 94.72 | 292 | | MCP-SVM | 98.73 | 13 |
|
查看原文
马永杰, 马芸婷, 陈佳辉. 结合卷积神经网络多层特征和支持向量机的车辆识别[J]. 激光与光电子学进展, 2019, 56(14): 141001. Yongjie Ma, Yunting Ma, Jiahui Chen. Vehicle Recognition Based on Multi-Layer Features of Convolutional Neural Network and Support Vector Machine[J]. Laser & Optoelectronics Progress, 2019, 56(14): 141001.
PDF全文
