Traffic Sign Recognition Method Based on Dual-Channel CNN

Aiming at traffic sign problem, the traditional LeNet-5 network structure has low accuracy of traffic sign recognition, slow identification speed and ignores natural factors such as weather. A convolutional network structure model with two-channel and multi-scale based on LeNet-5 improvement is prop...

Celý popis

Uložené v:
Podrobná bibliografia
Vydané v:2021 IEEE 15th International Conference on Electronic Measurement & Instruments (ICEMI) s. 364 - 369
Hlavní autori: Zhao, Zeyi, Zhou, Fuqiang, Wang, Shaohong, Xu, Xiaoli, Chao, Fenjin, Lin, Zhoupeng
Médium: Konferenčný príspevok..
Jazyk:English
Vydavateľské údaje: IEEE 29.10.2021
Predmet:
convolutional neural network
Convolutional neural networks
Feature extraction
Instruments
LeNet-5 network structure
Neural networks
Robustness
Stochastic processes
traffic sign recognition
Training
On-line prístup:Získať plný text
Tagy: Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
Popis
Shrnutí:Aiming at traffic sign problem, the traditional LeNet-5 network structure has low accuracy of traffic sign recognition, slow identification speed and ignores natural factors such as weather. A convolutional network structure model with two-channel and multi-scale based on LeNet-5 improvement is proposed by convolutional neural network technology. In the dual-channel structure, each channel contains two branching structure, and the number of convolution and image scale of each channel is different, making the feature extraction of different image scales richer. Secondly, the improved network structure greatly increases the number of convolutional kernels compared to the traditional LeNet-5 network structure. Finally, by changing the Sigmoid activation function to the ReLu activation function, changing the stochastic gradient descent algorithm to the Adam algorithm, and adding Dropout layers to prevent overfitting and setting the learning rate, thus increasing the traffic sign recognition rate. The recognition rate of the improved network is 98.6%, floating by 0.5%, relative to the traditional LeNet-5 network structure, the recognition rate increases by more than 15%, verifying that the improved network structure has a certain robustness.
DOI:10.1109/ICEMI52946.2021.9679593