目标边缘清晰化的图像风格迁移  

Image Style Transfer with Clear Target Edges

作　　者：沈瑜[1] 杨倩 苑玉彬 王霖 Shen Yu;Yang Qian;Yuan Yubin;Wang Lin(College of Electronics and Information Engineering,Lanzhou Jiaotong University,Lanzhou,Gansu730070,China)

机构地区：[1]兰州交通大学电子与信息工程学院,甘肃兰州730070

出　　处：《激光与光电子学进展》2021年第12期225-237,共13页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61861025,61663021,61761027,61669010);长江学者和创新团队发展计划(IRT16R36);兰州市人才创新创业项目(2018-RC-117);光电技术与智能控制教育部重点实验室(兰州交通大学)开放课题(KFKT2018-9)。

摘　　要：针对图像风格迁移时出现前后景边界模糊造成风格化图像主要目标模糊的问题,提出了目标边缘清晰化的图像风格迁移算法。通过将用于提取内容图像轮廓的深度抠图神经网络与风格迁移网络合并,形成透明遮罩约束风格迁移过程,凸显风格化图像中主要目标的轮廓;通过对迁移网络中最大池化层进行替换,保留图像的背景信息,细化风格化图像的整体结构;通过替换迁移网络中较大卷积核,减少网络模型参数,减少风格迁移计算量;通过对常规卷积层的归一化,实现相似风格迁移之间的参数共享,提升风格迁移速度。用VGG-19神经网络作为特征提取器对输入的内容图像和风格图像提取特征图,把输入图像到输出图像的变换约束在色彩空间局部仿射中,在输入图像RGB通道上合并目标遮罩,使得风格化图像的主要目标在遮罩约束中实现纹理合成。实验表明,与传统方法比较,该方法产生的迁移结果前后景边缘明显,内容结构保留较好,解决了风格化图像主要目标边缘模糊的问题。In the process of image style transfer,the main target of the stylized image is blurred due to the blurring of the foreground and background boundaries in image reconstruction.Image style transfer algorithm with clear target edges is proposed.The deep matting neural network used to extract outline of the content image is merged with the style transfer network to form a transparent mask to constrain the style transfer process,highlighting the outline of the main target of the stylized image.By replacing the max-pooling layer in the transfer network,more image background information is retained,and the overall structure of the stylized image is refined.The network model parameters are reduced by replacing the larger convolution kernel in the transfer network,and the calculation process of style transfer is simplified.Then,the normalization of the conventional-convolutional layer realizes parameter sharing between similar style transfers,improving the speed of style transfer.The VGG-19 neural network as a feature extractor extracts feature maps from the input content and style images,constrains the transformation from the input image to the output image in local affine of color space,and merges the target mask on the RGB channel of the input image so that the main goal of stylized images is to achieve texture synthesis in mask constraints.Experimental results show that compared with the traditional method,this method has obvious boundaries in the foreground and background of the stylized image,and the content structure is preserved well.

关 键 词：图像处理 风格迁移 神经网络 抠图算法 深度学习 结构约束 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]