基于自适应归一化的语义图像合成

Semantic Image Synthesis with Adaptive Normalization

本文提出了一种可以实现风格控制的自适应归一化。它是一个简单但有效的模块,应用于以分割掩膜为条件的生成对抗网络。以前的方法将风格图像作为输入,输入到深度网络中。本文方法通过在归一化层输入风格信息来学习参数,以此调节归一化层的激活。本文在两个数据集上进行实验,并展示了部分结果。结果表明,本文方法可以根据语义分割掩膜合成符合语义布局和视觉逼真度高的图像,并以同一的模型实现不同风格的转换。

This paper presents an adaptive normalization method which can realize style control. It is a simple but effective module, which is applied to generate countermeasure network under the condition of segmented mask. The traditional method takes the style image as the input and inputs it into the depth network. In this paper, the method learns parameters by inputting style information in the normalization layer, so as to adjust the activation of the normalization layer. In this paper, experiments are carried out on two data sets, and some results are shown. The results show that this method can synthesize images with high semantic layout and visual fidelity according to the semantic segmentation mask, and realize the transformation of different styles with the same model.