基于多维特征和语境信息融合的车牌检测方法

License Plate Detection Based on Multi-characteristic and Context Feature Fusion

为了提高交通监控视频中不同拍摄距离和拍摄角度下车牌检测的性能,提出了一种基于深度卷积神经网络,利用多维特征信息增强和语境信息融合优化车牌检测性能的算法。首先,在标注一块区域称为车牌上下文区域,结合车辆在图中的位置作为辅助车牌检测的语境信息。接着,为了提取出车牌区域和语境区域,对两阶段检测网络Faster R-CNN做出调整:选取VGG16中不同的卷积层输出分别融合成针对车辆区域,车牌上下文区域,车牌区域的多尺度融合特征图,使低层位置信息和高层语义信息得以互补,增强特征的表征能力,减小尺寸因素的影响。随后对检测到的车牌区域特征和语境区域特征进行融合,实现车牌检测的修正。最后,在RPN阶段,用旋转anchor替换矩形anchor来生成更加合适的预测框,解决真实场景中由观测角度引起的车牌旋转问题。基于多个基准车牌数据库的实验结果表明,文中提出的算法与现有算法相比,针对不同尺寸和不同角度的车牌具有更好的检测效果。

In order to improve the performance of license plate detection under different shooting distances and shooting angles in traffic surveillance video,an algorithm based on deep convolutional neural network with multi-dimensional feature information enhancement and context information fusion to optimize the performance of license plate detection is proposed.Firstly,we introduce a region called context-of-plate combined with vehicle location as the context information,exploiting the hidden correlation.And to extract local and contextual features,we make some modifications to the Faster-RCNN:selecting several layers with different shades of VGG16 for vehicle,context-of-plate and license plate to obtain multi-scale integrated feature maps,complementing location and semantic information,enhancing the representation of features and reducing scale interference.Then,the features of license plate region and the context region are fused to refine the license plate detection.Finally,in the RPN stage,the rectangular anchor is replaced with a rotating anchor to generate a more appropriate prediction box and solve the license plate rotation problem caused by the observation angle in the real scene.Experiments on benchmark datasets demonstrate that the proposed method shows better performance compared with existing methods under various shooting instance and observation angles.