Vehicle color recognition based on SqueezeNet(PDF)
长安大学学报（自然科学版）[ISSN:1006-6977/CN:61-1281/TN]
- Issue:
- 2020年4期
- Page:
- 109-116
- Research Field:
- 交通工程
- Publishing date:
Info
- Title:
- Vehicle color recognition based on SqueezeNet
- Author(s):
- LI Kunlun1; 2; WEI Zefa1; SONG Huansheng1; 3
- (1. Educational Technology and Network Center, Changan University, Xian 710064, Shaanxi, China;2. School of Electronic and Control Engineering, Changan University, Xian 710064, Shaanxi, China;3. School of Information Engineering, Changan University, Xian 710064, Shaanxi, China)
- Keywords:
- traffic engineering; convolutional neural network; intelligent transportation; vehicle color recognition; SqueezeNet
- PACS:
- -
- DOI:
- -
- Abstract:
- In order to accurate identification of vehicle colors which can assist existing license plate recognition system and enable accurate identification of vehicles in intelligent transportation system, existing solutions were analyzed, and the problems of the practical applications of convolutional neural network (CNN) were explored. A method of realizing vehicle color recognition by using light weight convolutional neural network (LCNN) was proposed. The parameter quantity of SqueezeNet is one fiftieth of that of AlexNet network structure, and it can also ensure the accuracy, which avoids the largescale calculation and high computer hardware requirements caused by the complex network structure and large parameter quantity, and therefore improves the portability of the model. The vehicle color recognition data set was selected as the research basis, and the data set was expanded and enhanced. Based on SqueezeNet, the impact of feature fusion on classification results was explored. The output feature map of fire7/concat and the output feature map of fire9/concat were fused. The results show that SqueezeNet has a training accuracy of 96.28% on the basis of guaranteed model size of 2.9 M and a single sheet test time of 15 ms, while that of AlexNet is 227 M, 24 ms and 96.18%. It can be seen that, under the premise of achieving the same accuracy, the model trained by the lightweight convolutional neural network SqueezeNet is more suitable to be transplanted to the development board such as fieldprogrammable aate array (FPGA), and the processing speed on the server is faster. Therefore, the lightweight convolutional neural network SqueezeNet, which is used to identify vehicle colors, can be better applied in the intelligent transportation system and solve, to a certain extent, current problems and drawbacks. 3 tabs, 7 figs, 23 refs.
Last Update: 2020-07-31