基于图像矫正与去噪的车牌识别算法RD-LPRNet

RD-LPRNet:A License Plate Recognition Algorithm Based on Image Rectification and Denoising

复杂场景下的车牌目标存在大倾角、远距离、光照不均等负面因素,极大地加剧了车牌识别的难度。针对上述问题,提出了一种新的车牌识别算法RD-LPRNet,该算法首先采用改进空间变换网络RNet预测仿射变换参数,对存在大倾角的车牌进行几何矫正;接着构造基于自适应注意力机制的去噪网络DNet,以真实-标准车牌对为训练数据,提高网络提取有效特征、剔除无效噪声信息的能力;最后基于LPRNet,改进其网络输入结构以丰富初始输入特征,提高训练收敛速度;添加深度分离卷积瓶颈结构,增强网络的特征提取能力;加入残差卷积替代特征拼接实现更高效的特征融合与传递。对CCPD、CLPD、PKUdata、OpenITS以及本文人工采集的车牌数据的测试结果表明,RD-LPRNet在复杂场景下表现出更高的识别性能。

In complex scenarios,license plate recognition is greatly challenged by negative factors such as large tilt angles,long distances,and uneven illumination of the license plate targets.To address these problems,a new license plate recognition algorithm called RD-LPRNet is proposed in this paper.Firstly,the algorithm uses an improved spatial transformation network(STN)to predict affine transformation parameters and perform geometric rectification on license plates with large tilt angles.Then,a denoising network with an adaptive attention mechanism is constructed,which improves the network's ability to extract useful features and eliminate noise information using real-standard license plate pairs as training data.Finally,based on LPRNet,the input structure is redesigned to enrich the initial input features and to improve the training convergence speed.Depth-wise separable convolution bottleneck structure is introduced to enhance the feature extraction ability.Residual connection instead of feature concatenation is used to achieve more efficient feature fusion and propagation.Experimental results on CCPD,CLPD,PKUdata,OpenITS,and a manually collected license plate dataset show that RD-LPRNet exhibits higher recognition performance in complex scenarios.