基于生成式对抗网络的发票图像超分辨率研究  

Research of Super-resolution Processing of Invoice Image Based on Generative Adversarial Network

作　　者：李新利[1] 邹昌铭 杨国田[1] 刘禾[1] Li Xinli;Zou Changming;Yang Guotian;Liu He(School of Control and Computer Engineering,North China Electric Power University,Beijing,102206,China)

机构地区：[1]华北电力大学控制与计算机工程学院,北京102206

出　　处：《系统仿真学报》2021年第6期1307-1314,共8页Journal of System Simulation

摘　　要：发票自动识别可有效提高财务工作效率。为避免低分辨率的发票图像影响自动识别的准确性,提出了一种用于对发票图像进行超分辨率处理的ESRGAN(Encoder Super-resolution Generative Adversarial Network)网络。ESRGAN网络是基于带条件的生成式对抗网络,设计了辅助编码器,引导网络生成更加真实的超分辨率图像。基于实际发票图像,将ESRGAN网络与常规图像处理、SRCNN(Super-resolution Convolutional Neural Networks)网络和SRGAN(Super-resolution Generative Adversarial Network)网络进行对比实验,并通过峰值信噪比(Peak Signal to Noise Ratio, PSNR)和结构相似性(Structural Similarity, SSIM)评价指标进行模型评价。实验结果表明基于ESRGAN超分辨率处理的图像在视觉效果和评价指标上均具有良好的效果。Automatic identification of invoices can effectively improve financial efficiency. But low-resolution invoice image reduces the accuracy of automatic identification, an ESRGAN(Encoder Super-resolution Generative Adversarial Network) network for super-resolution processing of invoice images is proposed. The ESRGAN network is based on a conditional generative adversarial network. An auxiliary encoder is designed to guide the network to generate a more realistic super-resolution image. Based on the actual invoice image, the ESRGAN network and the conventional image processing, SRCNN(Super-resolution Convolutional Neural Networks) network and SRGAN(Super-resolution Generative Adversarial Network) network. The model is evaluated through two evaluation indicators of peak signal-to-noise ratio(PSNR) and structural similarity(SSIM). The experimental results show that the images processed based on ESRGAN super-resolution are better on visual effects and evaluation indicators.

关 键 词：发票图像 超分辨率 生成式对抗网络 ESRGAN 评价指标 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]