基于空间转换网络的端到端车牌检测与识别

End-to-end license plate detection and recognition based on spatial transformer network

在复杂场景中,许多现有的车牌检测和识别方面的研究方法存在数据集单一且有限、算法复杂等问题。因此提出了一个端到端的统一网络:残差-空间变换-连接时序分类融合的车辆号牌检测识别网络(LPDR-RSCNet)。该网络结合残差神经网络、空间变压器网络和连接主义者时间分类,联合训练检测和识别模块,以减少中间错误积累。通过在残差神经网络提取特征过程中引入空间变换网络,使特征提取器具有平移不变性、旋转不变性和缩放不变性;在分类器引入连接时序分类,可以自动识别图片标签和特征之间的关系。同时,还可以适应可变长度序列的识别。在中国城市停车场数据集(CCPD)上进行了比较实验,CCPD是一个大规模、多样的中文车牌数据集。实验证明LPDR-RSCNet模型在实际应用中可实现98.8%的识别精度和34 fps的速度,并且相较于YOLO9000、Faster-RCNN、SSD300,具有更好的检测准确度,可满足智能交通系统中对移动车辆实时车牌检测和识别的要求。

Many existing approaches perform poorly in the detection and recognition of license plates in complex scenes.So the LPDR-RSCNet-an end-to-end and unified network for the detection and recognition of license plates in a single forward pass was proposed.In this network,the detection and recognition modules are jointly trained to reduce intermediate error accumulation with residual neural network,spatial transformer network and connectionist temporal classification.By introducing spatial transformation network into the residual neural network feature extraction process,the extractor has translation invariance,rotation invariance and scaling invariance;by introducing connection time series classification into classifier,the relationship between image label and feature can be recognized automatically.At the same time,the network can also be adapted to the recognition of variable length sequences.The experiments are conducted on CCPD which is a large-scale and diverse Chinese license plate dataset.Compared with YOLO9000,Fast R-CNN and SSD300,the LPDR-RSCNet model has been proved to be more effective,greatly improves detection accuracy of license plates in complex field scene performance.In practical applications,this model achieves 98.8%recognition accuracy and 34 fps speed.This model can meet the requirements of real-time license plate detection and recognition of moving vehicles in Intelligent Traffic System.