基于卷积神经网络的图像局部风格迁移

Image Localized Style Transfer Based on Convolutional Neural Network

图像风格迁移是计算机图形学和计算机视觉的一个研究热点。针对现有的图像风格迁移方法中难以对内容图局部区域进行风格迁移的难点,提出了一种基于卷积神经网络的图像局部风格迁移框架。首先,根据输入的内容图和风格图,利用图像风格迁移网络生成全局风格迁移图;然后,利用图像语义分割网络,通过自动语义分割生成的掩码确定图像前景区域与背景区域;最后,利用掩码图确定风格迁移区域并融合未迁移区域得到图像局部风格迁移结果,同时提出一种基于曼哈顿距离的图像融合算法以优化局部风格迁移对象与未迁移区域之间边界的衔接和平滑过渡。该框架综合考虑了目标区域和边界带的像素值、位置等细节信息,在3个公开的图像数据集上进行实验,结果表明该方法能够高效、快速并自然地实现输入内容图的局部风格迁移,生成艺术性与真实性和谐并存的视觉效果。

Image style transfer is a research hot topic in computer graphics and computer vision.Aiming at the difficulty in the style transfer of the local area of the content image in the existing image style transfer method,this paper proposed a localized image transfer framework based on convolutional neural network.First,according to the input content image and style image,the image style transfer network is used to generate the whole style transferred image.Then,the image foreground and the background area are determined by the mask generated by automatic semantic segmentation.Finally,according to style transfer result of the foreground or the background region,an image fusion algorithm based on Manhattan distance is proposed to optimize the convergence and smooth transition between the stylized object and the original area.The framework comprehensively considers the pixel values and positions of the target area and the boundary band,and experiments on three public image datasets demonstrate that the method can efficiently,quickly and naturally implement local style transfer of input content maps,and produce visual effects that are both artistic and authentic.