激光与光电子学进展, 2021, 58 (2): 0210013, 网络出版: 2021-01-08
共享型轻量级卷积神经网络实现车辆外观识别 下载: 1040次
Vehicle Appearance Recognition Using Shared Lightweight Convolutional Neural Networks
图像处理 共享型轻量级卷积神经网络 颜色特征 型号特征 改进的SqueezeNet 车辆外观识别 image processing shared lightweight convolutional neural network color feature type feature improved SqueezeNet vehicle appearance recognition
摘要
提出一种共享型轻量级卷积神经网络(CNN),用于自动识别车辆颜色和型号。基础网络采用改进的SqueezeNet,在训练集上比较具有不同“瘦身”程度的SqueezeNet的分类性能。讨论完全共享型网络、部分共享型网络及无共享型网络的特征。实验结果表明,完全共享型轻量级CNN在减少参数量的同时实现了对车辆外观多属性的高精度识别。在开放数据集Opendata_VRID上进行实验,车辆颜色和车型识别的准确率分别达98.5%和99.1%。在一台无GPU配置的个人计算机上,单张图片的识别时间仅为4.42 ms。共享型轻量级CNN大大减少了时间和空间成本,更有利于在资源有限的环境中进行部署。
Abstract
In this study, we propose a shared lightweight convolutional neural network (CNN) to automatically identify vehicle colors and types. In the basic network, an improved SqueezeNet is employed. Further, we compare the classification performances of different “slimming” SqueezeNets on the training set. In addition, the characteristics of the fully shared, partly shared, and no-shared networks are discussed. Experimental results indicate that the fully shared lightweight CNN not only reduces the number of parameters but also realizes high-precision recognition of the multiple attributes associated with the appearance of vehicles. Subsequently, an experiment was conducted on the Opendata_VRID dataset. The accuracy of vehicle color and type recognition is 98.5% and 99.1%, respectively. A single picture can be recognized on a personal computer without GPU in only 4.42 ms. Thus, the shared lightweight CNN considerably reduces time and space consumption and is more conducive for deployment in resource-constrained systems.
康晴, 赵红东, 杨东旭. 共享型轻量级卷积神经网络实现车辆外观识别[J]. 激光与光电子学进展, 2021, 58(2): 0210013. Qing Kang, Hongdong Zhao, Dongxu Yang. Vehicle Appearance Recognition Using Shared Lightweight Convolutional Neural Networks[J]. Laser & Optoelectronics Progress, 2021, 58(2): 0210013.