一种改进卷积循环神经网络的复杂场景下的车牌识别模型

An Improved Convolutional Recurrent Neural Network for License Plate Recognition Model in Complex Scenes

摘要: 识别自然场景图像中的汽车牌照是一项重要而又具有挑战性的任务。许多现有方法对于在固定场景下收集的牌照表现良好，但它们的性能在诸如车牌角度倾斜、光照强度过亮或过暗、图片模糊等复杂的环境中显著下降。本文提出了一种改进的卷积循环神经网络车牌识别模型，在网络中加入幻影模块(Ghost Block)和卷积块注意模块(Convolutional Block Attention Module, CBAM)，能够提高车牌字符特征提取的丰富程度的同时在通道和空间方向上对车牌字符特征进行加权，提高模型对车牌字符识别的准确率。最后通过实验验证了本文提出的模型的有效性。

Abstract: Recognizing car license plates in images of natural scenes is an important and challenging task. Many existing methods perform well for license plates collected in fixed scenes, but their perfor-mance degrades significantly in complex environments such as tilted license plate angles, too bright or too dark light intensities, and blurred images. In this paper, we propose an improved convolu-tional recurrent neural network license plate recognition model by adding Ghost Block and Convo-lutional Block Attention Module (CBAM) module to the network, which can improve the richness of license plate character feature extraction while weighting the license plate character features in channel and spatial direction to improve the accuracy of the model for license plate character recognition. Finally, the effectiveness of the model proposed in this paper is verified by experi-mental evidence.