Classifier for Recognition of Fine-Grained Vehicle Models under Complex Background

Jie Zhang; Hongdong Zhao; Yuhai Li; Miao Yan; Zetong Zhao

doi:10.3788/LOP56.041501

Laser & Optoelectronics Progress, Volume. 56, Issue 4, 041501(2019)

Classifier for Recognition of Fine-Grained Vehicle Models under Complex Background

Jie Zhang^1,2, Hongdong Zhao^1,2, Yuhai Li^2、*, Miao Yan¹, and Zetong Zhao¹

Author Affiliations

¹ School of Electronic and Information Engineering, Hebei University of Technology, Tianjin 300401, China

² Science and Technology Electro-Optical Information Security Control Laboratory, Tianjin 300308, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(9)

Fig. 1. Flow chart of Softmax-SVM classifier

Download full size

Fig. 2. Structural diagram of Softmax-SVM classifier based on DCNN

Download full size

Fig. 3. Partial samples of 27 types of vehicle models

Download full size

Fig. 4. DCNN performance versus training under different loss functions. (a) Training accuracy; (b) test accuracy; (c) training loss; (d) test loss

Download full size

Table 1. Recognition accuracies of DCNN training for 350 times under different loss functions
View table
Table 1. Recognition accuracies of DCNN training for 350 times under different loss functions
Loss function Accuracy /%
Training set Test set
Square-loss 80.0 76.58
Cross-entropy-loss 99.8 95.51
Exp-loss 56.0 57.28

Table 2. Losses of DCNN training for 350 times under different loss functions
View table
Table 2. Losses of DCNN training for 350 times under different loss functions
Loss function Loss/arb. units
Training set Test set
Square-loss 0.3729 0.3886
Cross-entropy-loss 0.0810 0.2144
Exp-loss 0.6354 0.6448

Table 3. Test accuracies of different classifier models
View table
Table 3. Test accuracies of different classifier models
Feature extraction Classifier Accuracy /%
HOG SVM 87.59
SURF Bag for word 46.00
DCNN Softmax 95.51
DCNN Softmax-SVM 97.78

Table 4. Time for training and recognition of all test samples by different classifiers
View table
Table 4. Time for training and recognition of all test samples by different classifiers
Featureextraction Classifier Time /s
Training set Test set
HOG SVM 195.212 2.127
SURF Bag 279.534 8.869
DCNN Softmax 312.386 2.441
DCNN Softmax-SVM 239.690 0.759

Table 5. Softmax-SVM performance in extracting features of different layers of DCNN to train SVM
View table
Table 5. Softmax-SVM performance in extracting features of different layers of DCNN to train SVM
Layer Accuracy /% Loss Time /s
FC2 97.03 1.0024 0.4609
ReLU6 97.40 1.0020 0.7639
FC1 97.78 1.0018 0.7591
ReLU5 95.54 1.0027 4.5576

Tools

Get Citation

Copy Citation Text

Jie Zhang, Hongdong Zhao, Yuhai Li, Miao Yan, Zetong Zhao. Classifier for Recognition of Fine-Grained Vehicle Models under Complex Background[J]. Laser & Optoelectronics Progress, 2019, 56(4): 041501

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Machine Vision

Received: Aug. 16, 2018

Accepted: Sep. 4, 2018

Published Online: Jul. 31, 2019

The Author Email: Yuhai Li (695263295@qq.com)

DOI:10.3788/LOP56.041501

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Recognition accuracies of DCNN training for 350 times under different loss functions

Table 1. Recognition accuracies of DCNN training for 350 times under different loss functions

Table 2. Losses of DCNN training for 350 times under different loss functions

Table 2. Losses of DCNN training for 350 times under different loss functions

Table 3. Test accuracies of different classifier models

Table 3. Test accuracies of different classifier models

Table 4. Time for training and recognition of all test samples by different classifiers

Table 4. Time for training and recognition of all test samples by different classifiers

Table 5. Softmax-SVM performance in extracting features of different layers of DCNN to train SVM

Table 5. Softmax-SVM performance in extracting features of different layers of DCNN to train SVM