基于可变形卷积和自适应二维位置编码的鲁棒车牌识别方法

车牌识别是智能交通系统中的关键步骤，为提高在非可控和复杂场景下车牌的识别精度，提出了一种鲁棒车牌识别方法，该方法主要包括车牌检测和车牌字符识别2个核心部分。首先，采用YOLOv5网络实现车牌的检测；其次，基于递归卷积神经网络框架，提出了一种基于可变形卷积和自适应二维位置编码(A2DPE)的车牌字符识别方法。该方法针对车牌大小、倾斜角度和光照条件等动态变化的特点，采用了可变形卷积来更好地提取车牌字符的特征，并引入了A2DPE模块，根据输入自适应地获取车牌字符位置编码信息。最后，利用双向长短期记忆网络进行车牌字符的识别，无须分割车牌字符，可以实现不同长度车牌字符的准确识别。在自建数据集LPdata与公开数据集CLPD上的实验结果表明，与现有方法相比，该方法能够以较低的模型复杂度达到较高的准确率。

Robust License Plate Recognition Based on Deformable Convolution and Adaptive 2D Positional Encoding

License plate recognition is a key step in intelligent transportation systems.In order to improve the recognition accuracy of license plates in uncontrollable and complex scenarios,a robust license plate recognition method is proposed,which mainly consists of two core parts:license plate detection and license plate character recognition.Firstly,YOLOv5 network is used to achieve the detection of license plates.Secondly,a license plate character recognition method based on deformable convolution and adaptive 2D positional encoding (A2DPE) is proposed based on convolutional recurrent neural network framework.The method employs deformable convolution to better extract the features of license plate characters for the dynamic changes of license plate size,tilt angle,and lighting conditions,and introduces A2DPE module to adaptively obtain the license plate character position encoding information based on the input.Finally,bidirectional long-short term memory network is used for license plate character recognition,which can achieve accurate recognition of license plate characters of different lengths without license plate character segmentation.The experimental results on the self-built dataset LPdata and the CLPD public dataset show that the method can achieve higher accuracy with lower model complexity compared with existing methods.