基于多分支卷积神经网络的车辆图像比对方法

为了更有针对性地从车辆图像的不同区域提取出独特的图像特征,提出基于多分支卷积神经网络的车辆图像比对方法.首先,根据车牌定位结果获取比对车辆的车脸图像,并根据车脸图像的纹理丰富度将车脸图像划分为多个图像块;其次,使用多分支卷积神经网络分别提取各车脸图像块的深度特征;最后,通过计算车脸图像深度特征的相似度判定比对的车辆图像是否属于同种车型.实验表明,提出的方法能够提取有效的车辆图像各区域的深度特征,获得良好的车辆比对准确率,可用于套牌车辆识别.

Vehicle Image Verification Based on Multi-branch Convolutional Neural Network

To more targeted extract special image features from different regions of the vehicle images, this paper proposes the vehicle image verification method based on multi-branch convolutional neural network. Firstly, obtain vehicle face by the location of vehicle plate, and segment entire vehicle face into many parts by the richness of texture of vehicle images. Secondly, use multi-branch convolutional neural network to extract deep features of vehicle face image. Finally, vertify the input vehicle faces by the similarity between their deep features. Experiments demonstrate that the proposed method can extract efficient deep features of vehicle image, and achieve higher verification accuracy, can be used for fake-licensed car identification.