基于CRNN的车牌识别方法

CRNN-Based License Plate Recognition Method

车牌识别是道路交通、智慧城市建设的重要组成部分,传统的车牌识别需要先检测出车牌位置,然后通过像素映射等方法分割出单个字符,最后利用模板匹配等方法进行识别。整个过程不仅速度慢,而且操作繁琐,分割或识别的效果也很难令人满意。本文基于YOLOv4-tiny和卷积循环神经网络(Convolution Recurrent Neural Network, CRNN)提出了一种端到端的方法。该方法利用注意力机制与YOLO4-tiny的融合,有效且快速的检测车牌位置,然后利用空间变换网络(Spatial Transformer Networks, STN)、残差学习(Residual Learning)以及注意力机制(Attention)与CRNN的融合高效的识别车牌信息。本文使用平均精度(Average Precision, AP)和识别准确率(Accuracy)作为检测和识别结果的主要评估指标。实验结果表明,车牌检测模型在交并比(Intersection-over-Union, IoU)为0.5的前提下AP值达到了93.60%,并且识别模型在蓝牌、绿牌的混合车牌下达到了92.15%左右的识别准确率。该方法相比于之前的车牌识别模型,不但识别准确率更高,而且能够直接通过该模型识别混合车牌,大大减少了现实情况下车牌识别的复杂度。

License plate recognition is an important part of road traffic and smart city construction. Traditional license plate recognition needs to detect the position of the license plate first, then segment a single character by pixel mapping, and finally use template matching and other methods for recognition. The whole process is not only slow, but also cumbersome to operate, and the effect of segmentation or recognition is difficult to be satisfied. This paper proposes an end-to-end method based on YOLOv4-tiny and Convolution Recurrent Neural Network (CRNN). This method uses the fusion of the attention mechanism and YOLO4-tiny to effectively and quickly detect the position of the license plate, and then uses the spatial transformation network (STN), residual learning, attention mechanism and CRNN to efficiently recognition of license plate information. This article uses Average Precision (AP) and Recognition Accuracy as the main evaluation indicators for detection and recognition results. The experimental results show that the AP value of the license plate detection model reaches 93.60% under the premise that the Intersection-over-Union (IoU) is 0.5, and the recognition accuracy reaches about 92.15% under the mixed license plate of blue and green plates. Compared with the previous license plate recognition model, this method not only has higher recognition accuracy, but also can directly recognize mixed license plates through the model, which greatly reduces the complexity of license plate recognition in real situations.