Abstract:

The fast arbitrary style transfer based on meta-networks has attracted great attention and high praise. However, visible gray blocks of stylistic incongruity often appear in stylized result image. The hue of stylized result image is often not consistent with target style image, which severely affects qualities of final transfer results. This paper proposes an improvement strategy. Gram matrix is used as a style statistic for meta network input and loss computing. By integrating the average pooling operation of Gram matrix and the grouped full connection strategy of meta-network, this paper effectively avoids the problem of too large network parameters brought by traditional Gram matrix. Experimental results show that the method can not only effectively eliminate the incongruous style block problem, but also achieve better visual effect than the original method in texture and color layout. Through theoretical analysis and experimental evidence, this paper confirms the superiority of using Gram matrix as style loss and feature statistics on convergence and visual effect.

Key words: meta network, style transfer, Gram matrix, arbitrary style, neural network

摘要：

基于元网络的任意风格快速迁移方法得到业界的高度关注和评价。然而，该模型的结果图中经常出现灰色风格不协调的像素缺块，颜色色调与目标风格图不一致，严重影响了迁移质量。提出了该方法的改进策略。使用Gram矩阵作为风格统计量，用于元网络信息输入和计算网络训练损失函数。同时，综合Gram矩阵平均池化操作和元网络分组全连接策略，有效避免了传统Gram矩阵带来网络参数整体过大的问题。实验结果显示，该方法不仅有效去除了不协调风格缺块问题，而且在纹理和颜色布局上较原方法取得了更好的视觉效果。通过理论分析、实验佐证，在算法收敛性和视觉效果方面，进一步确认了采用Gram矩阵作为风格损失和特征统计量的优越性。

关键词: 元网络, 风格迁移, Gram矩阵, 任意风格, 神经网络