基于多视角序列图像的高光去除CycleGAN网络  

Cyclegan Network for Highlight Removal Based on Multi-view Sequence Images

作　　者：郭圣逸 李丽 沈彬[1,2,3] 陈常念 胡新荣 GUO Shengyi;LI Li;SHEN Bin;CHEN Changnian;HU Xinrong(Hubei Provincial Engineering Research Center for Intelligent Textile and Fashion,Wuhan 430200,China;Engineering Research Center of Hubei Province for Clothing Information,Wuhan 430200,China;School of Computer Science and Artificial Intelligence,Wuhan Textile University,Wuhan 430200,China)

机构地区：[1]纺织服装智能化湖北省工程研究中心,湖北武汉430200 [2]湖北省服装信息化工程技术研究中心,湖北武汉430200 [3]武汉纺织大学计算机与人工智能学院,湖北武汉430200

出　　处：《郑州大学学报（理学版）》2023年第5期11-16,共6页Journal of Zhengzhou University:Natural Science Edition

基　　金：国家自然科学基金青年项目(61901308)。

摘　　要：光线照射镜面物体产生的镜面反射使得采集的图像产生高光现象,高光会影响很多视觉任务的精度。针对图片的去高光问题,在经典无监督学习CycleGAN的框架下提出了一种端到端的分层网络,该模型的输入为已标定的序列高光图像,输出为去除高光的图像。为了获取成对数据集以训练网络,使用可微分渲染器生成视角、光照可控镜面反射-漫反射成对合成数据集。无监督CycleGAN图像风格迁移网络作用于输入图像时,仅使用小批量的背景图片,即可将图像分解为前景与背景,图像风格迁移网络仅作用于前景,进一步提高了图像转换的精度。实验结果表明,该方法可有效去除高光。The specular reflection produced by the light shining on the specular object caused the captured image to produce highlights,and the highlight would affect the accuracy of many visual tasks.Aiming at the problem of de-highlighting of images,an end-to-end layered network was proposed in the framework of the classic unsupervised learning CycleGAN.The input of the model was the calibrated sequence highlight images,and the output was the image with the highlights removed.To obtain paired datasets to train the network,a differentiable neural renderer was used to generate a pairwise synthetic dataset of viewing angle,lighting controllable specular-diffuse reflections.When the unsupervised CycleGAN image style transfered network acted on the input image,a small batch of background pictures was used to decompose the image into foreground and background.The image style transfered network only acted on the foreground,so the accuracy of image conversion was further improved.The experimental results showed that the method was effective for accurate highlight removal.

关 键 词：高光去除 卷积神经网络 CycleGAN 无监督 可微分渲染 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]