基于深度残差网络和注意力机制的特殊车牌识别  

Deep residual network and attention mechanism for special license plate recognition

作　　者：王昊 陈黎[1,2] WANG Hao;CHEN Li(School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China;Hubei Province Key Laboratory of Intelligent Information Processing and Real-Time Industrial System,Wuhan University of Science and Technology,Wuhan 430065,China)

机构地区：[1]武汉科技大学计算机科学与技术学院,湖北武汉430065 [2]武汉科技大学湖北省智能信息处理与实时工业系统重点实验室,湖北武汉430065

出　　处：《计算机工程与设计》2024年第1期291-298,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(61773297)。

摘　　要：为解决现有车牌识别算法在面对旋转倾斜车牌以及双行车牌图像时识别精度偏低的问题,提出一种基于深度残差网络和注意力机制的特殊车牌识别算法。优化深度残差网络结构,使模型更好提取低分辨率车牌图像的特征;取消对特征图平均池化操作,在保留图像全局特征的前提下,将多维特征化为特征序列;引入注意力机制对特征序列并行解码,加快模型推理速度,提升特殊车牌的识别精度。实验结果表明,与现有的文字识别模型CRNN、DAN、ASTER对比,在公开车牌数据集CCPD上取得了更高的准确率,验证了模型的有效性。To address the problem of low recognition accuracy of existing license plate recognition algorithms facing rotating and tilting license plates and two-line license plate images,a special license plate recognition algorithm based on depth residual network and attention mechanism was proposed.The structure of the depth residual network was optimized to enable the model to better extract image features when facing low-resolution license plate images.The averaging pooling operation of feature maps was eliminated and the multidimensional feature maps were transformed into feature sequences while preserving the global features of images,and the attention mechanism was introduced to decode the feature sequences in parallel to speed up the model inference and improve the recognition accuracy of special license plates.After experiments,it is shown that comparing with the existing text recognition models CRNN,DAN,ASTER,higher accuracy is achieved on the public license plate dataset CCPD,which verifies the effectiveness of the model.

关 键 词：车牌识别 文字识别 多头注意力 自注意力机制 卷积神经网络 循环神经网络 残差网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]