Small vehicle classification in the wild using generative adversarial network
- PDF / 1,483,640 Bytes
- 11 Pages / 595.276 x 790.866 pts Page_size
- 76 Downloads / 176 Views
(0123456789().,-volV)(0123456789(). ,- volV)
ORIGINAL ARTICLE
Small vehicle classification in the wild using generative adversarial network Xu Wang1 • Xiaoming Chen1 • Yanping Wang1 Received: 8 January 2020 / Accepted: 2 September 2020 Ó Springer-Verlag London Ltd., part of Springer Nature 2020
Abstract With the popularization of intelligent transportation system, the demand for vision-based algorithms and performance becomes more and severe. Vehicle detection techniques have made great strides in the past decades; however, there are still some challenges, such as the classification of tiny vehicles. The images of distant vehicles are generally blurred and lack detailed information due to their low resolutions. To solve this problem, we propose a novel method to generate highresolution (HR) images from fuzzy images by employing a generative adversarial network (GAN). In addition, the dataset used for training standard GAN is generally constructed by down-sampling from the neutral HR images. Unfortunately, the effect of reconstruction is more modest. To cope with this trouble, we first construct our dataset by using three fuzzy kernels. Then, the exposure of the low-resolution (LR) image is adjusted randomly. Furthermore, a hybrid objective function is designed to guide the model to restore image details. The experimental results on the KITTI data set verify the effectiveness of our method for tiny vehicle classification. Keywords Vehicle classification GAN Image reconstruction Convolutional neural networks
1 Introduction Vehicle classification from videos and images is a crucial and basic problem in computer vision field, considering it is an essential precondition for intelligent traffic control, autonomous driving, and traffic surveillance. Great progress has been made in vehicle classification with the passage of time. Psyllos et al. [23] proposed an algorithm to identify vehicle’s manufacturer and model by capturing car trademarks. However, sometimes the trademark does not appear in car image because of different camera angles. On the contrary, Kafai et al. [17] extracted the feature set
& Xiaoming Chen [email protected] Xu Wang [email protected] Yanping Wang [email protected] 1
Department of Automatic Control, College of Automation Engineering, Nanjing University of Aeronautics and Astronautics, 29 Jiangjun Avenue, Jiangning District, Nanjing, Jiangsu Province, China
of taillights and vehicle dimensions. It is processed by a hybrid dynamic Bayesian network for classification. Shuang et al. [25] tried to use texture features for vehicle recognition and proposed a classification model based on feature encoding combined with CNN. It can achieve encouraging results, even though under the condition of little training samples. It is noteworthy that deep convolutional neural networks (CNNs) has been applied to vehicle detection together with assorted other object detection generally in the last few years because of its excellent performance. However, the classification results of the tiny vehicles o
Data Loading...
