Abstract:

Traditional grayscale image coloring methods have problems such as color distortion and poor effect, and have been gradually replaced by deep learning methods. At present, the portrait coloring method based on deep learning mainly has the problem of portrait miscoloring in complex background. In view of the above problems, this paper proposes a portrait coloring method for joint consistent cyclic generative adversarial network. The method uses a joint consistency loss training model based on the cycle generative adversarial network. The generative network uses the U-network structure （UNet） to improve the model generative network and enhance the integrity of the generated image information. The feature extraction method of multi-feature fusion is introduced in the adversarial network, and the feature expression of the feature is enhanced. Finally, through the comparison experiments in the self-built CASIA-PlusColors high-quality portrait dataset, it is verified that the proposed method has a better effect on portrait coloring in complex background.

Key words: portrait coloring, joint consistent cycle generative adversarial network, deep learning, feature fusion

摘要：

传统灰度图像着色方法存在颜色失真、效果不佳等问题，已逐渐被深度学习方法取代。目前基于深度学习的人像着色方法主要存在复杂背景下误着色的问题。针对上述问题，提出了联合一致循环生成对抗网络的人像着色方法。该方法在循环生成对抗网络的基础上，采用联合的一致性损失训练模型；生成网络采用U型网络结构（UNet）进行改进，以提高生成图像信息的完整性；判别网络中引入多特征融合的特征提取方式，增强特征对图像的细节表达。最后通过在自建的CASIA-PlusColors高质量人像数据集中的对比实验，验证了该方法对复杂背景中的人像着色有着更好的效果。

关键词: 人像着色, 联合一致循环生成对抗网络, 深度学习, 特征融合