基于超分辨率和深度神经网络的车型识别

车型识别在视频监控系统中起着关键作用,文中利用深度神经网络和超分辨率来实现交通监控中的车型识别。利用深度卷积神经网络CaffeNet,并采用先进的深度学习框架CAFFE和具有强大计算能力的GPU来完成对车辆的车型识别。在图像预处理阶段,采用一种基于深度学习和稀疏表示的图像超分辨率(SR)重构算法,来增强图像的细节信息。其中首先基于深度学习模型自编码器,提出一种改进模型非负稀疏去噪自编码器(Nonnegative Sparse Denoising Auto-Encoders,NSDAE)来实现字典的联合学习,然后基于稀疏表示实现车辆图像的超分辨率重构。经实验验证,在加入超分辨率处理之后,车型识别效果在精确度上得到了明显的提升。

Vehicle Recognition Based on Super-resolution and Deep Neural Networks

Vehicle recognition plays a key role in traffic video surveillance system.In this paper,deep neural network and super-resolution were uses to realize vehicle recognition in traffic surveillance.It uses deep convolution neural network CaffeNet to complete vehicle recognition with advanced deep learning framework CAFFE and computationally powerful GPU.In the image preprocessing stage,an image super-resolution reconstruction algorithm based on deep learning and sparse representation was used to enhance the detail information of the image.First,based on auto-encoders,an improved model of nonnegative sparse denoising auto-encoders (NSDAE) was proposed to realize the dictio-nary joint learning.Then,the sparse representation was used to realize super-resolution reconstruction of vehicle image.Experimental results show that the accuracy of vehicle recognition is improved obviously after adding the super resolution processing.