半监督塑封烟盒退化图像修复算法  

Semi-supervised image restoration algorithm for degraded plastic-sealed cigarette packs

作　　者：石彬 成苗 张绍兵 何莲 SHI Bin;CHENG Miao;ZHANG Shaobin;HE Lian(Chengdu Institute of Computer Application,Chinese Academy of Sciences,Chengdu Sichuan 610213,China;School of Computer Science and Technology,University of Chinese Academy of Sciences,Beijing 100049,China;Shenzhen CBPM-KEXIN Banking Technology Company Limited,Shenzhen Guangdong 518206,China)

机构地区：[1]中国科学院成都计算机应用研究所,成都610213 [2]中国科学院大学计算机科学与技术学院,北京100049 [3]深圳市中钞科信金融科技有限公司,广东深圳518206

出　　处：《计算机应用》2023年第S02期238-243,共6页journal of Computer Applications

摘　　要：针对塑封烟盒底纹修复没有成对数据且无监督域转换算法无法处理退化图像的问题,为了去除薄膜并完成底纹信息的修复,提出一种半监督塑封烟盒退化图像修复算法。首先,由薄膜图像生成子网络提取真实塑封烟盒的薄膜信息;然后,利用生成的薄膜图片和真实的底纹图片以及随机掩码(mask)合成数据集;最后,使用合成数据集学习有薄膜到无薄膜两个域之间的转换。在烟盒数据集上的实验结果表明,视觉定性比较,所提算法更好地修复了受薄膜影响而退化的底纹信息;数据定量比较,所提算法的弗雷歇初始距离(FID)比CycleGAN、基于对比学习的非成对图像翻译网络(CUT)、基于双重对比学习的无监督图像翻译网络DCLGAN分别降低了14.42%、6.85%和3.00%,有利于提高图片质量和方便后续检测,且单张样本平均推理耗时为13.73 ms,能够满足工业生产的实时要求。Aiming at the problems that there is no paired data for the shading restoration of plastic-sealed cigarette packs and the unsupervised domain transformation algorithm cannot deal with the degraded images,in order to remove the film and complete the restoration of the shading information,a semi-supervised image restoration algorithm for the degraded image of plastic-sealed cigarette pack was proposed.Firstly,the film information of the real plastic-sealed cigarette pack was extracted by the film image generation sub-network;then,the generated film image and the real shading image and random mask were used to synthesize the data set;finally,the synthetic data set was used to learn the transformation relationship between the two domains with film and without film.Experiments on the cigarette pack data set show that the proposed algorithm better restores the shading information degraded by the film in terms of visual qualitative comparison;in the quantitative comparison of data,compared with the algorithms such as unpaired image-to-image translation using CycleGAN,Contrastive learning for Unpaired image-to-image Translation(CUT),Dual Contrastive Learning for unsupervised image-toimage translation network DCLGAN,Fréchet Initial Distance(FID)is reduced by 14.42%,6.85%,and 3.00%respectively,which is conducive to improve the image quality and facilitate subsequent detection,and the average inference time of a single sample is 13.73 ms,which can meet the real-time requirements of industrial production.

关 键 词：跨域图像转化 半监督 图像修复 数据增广 深度学习 塑封烟盒 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]