基于Cycleflow-GAN的车牌图像生成方法  

License Plate Image Generation Method Based on Cycleflow-GAN

作　　者：张森 杨雨城 姚广芬 Zhang Sen;Yang Yucheng;Yao Guangfen(School of Rail Transportation,Wuyi University,Jiangmen Guangdong 529020,China)

机构地区：[1]五邑大学轨道交通学院,广东江门529020

出　　处：《山西电子技术》2024年第6期32-34,37,共4页Shanxi Electronic Technology

摘　　要：随着基于深度学习的方法在智能交通系统中的快速发展,车牌字符识别也发挥着至关重要的作用。然而,车牌图像样本不足会导致识别模型的性能不佳。基于此,提出了一种基于Cycleflow-GAN的图像生成方法,生成网络通过一系列结构简单的可逆变换函数构建,其可解释性强。同时优化了损失函数的复杂度,并且采用最小二乘损失,使模型的训练更加稳定。结果表明提高了生成图像的质量,生成的多样化车牌图可满足车牌字符识别大样本的训练。With the rapid development of deep learning-based methods in intelligent transportation systems,license plate character recognition also plays a crucial role.However,insufficient license plate image samples can lead to poor performance of the recognition model.In this work,an image generation method based on Cycleflow-GAN is proposed,where the generative network is constructed by a series of invertible transformation functions with simple structure and its interpretability is strong.The complexity of the loss function is also optimized and the least squares loss is used to make the training of the model more stable.The results show that the quality of the generated images is improved,and the generated diverse license plate maps can meet the training of large samples for license plate character recognition.

关 键 词：数据增强 流模型 图像生成 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]