【摘要】 中国传统的工笔花卉画中人工渲染上色过程繁复、技巧性较高,现有的线稿图自动上色算法难以生成自然合理的渐变色效果.文中基于条件生成对抗网络,提出注意力引导的交互式工笔花卉上色算法,自动完成花卉白描线稿到工笔花卉色图的仿真生成.首先设计刻画花朵颜色类别和布局的花色注意力图,可在训练阶段指导网络进行重要颜色特征的学习,在应用阶段作为用户与系统的交互手段,完成色彩设计.其次,在网络结构设计方面,构建并训练针对花色注意力图的局部颜色编码子网络,将注意力图的编码信息作为仿射参数,引入生成器各层的条件归一化过程中,实现生成网络全域对颜色的学习和控制.考虑到深度特征偏重刻画全局语义信息,可能损失反映线条特征的局部高频信息,在生成器网络中引入跨层连接的结构,加强线条特征的学习.实验表明,文中算法可以较好地将花卉白描线稿渲染成工笔花卉的色图,生成的图像符合真实工笔花卉画的颜色分布和特点,具有较好的艺术真实感和观赏性.更多还原

【Abstract】 The process of manually rendering traditional Chinese meticulous flower painting is complicated and highly skilled. The existing automatic line drawing colorization is difficult to generate natural and reasonable gradient effect. On the basis of condition generative adversarial network(CGAN), an interactive meticulous flower coloring algorithm via attention guidance is proposed to accomplish the colorization of meticulous flowers from line drawing. A color attention map depicting the color category and layout of flowers is designed to guide the proposed network to learn important color features in the training stage. The color attention map is considered as the means of interaction between the user and the system for color design in the application stage. In the network structure design, a local color-coding sub-network is constructed and trained to encode the flower color attention map. The encoded information is introduced into the conditional normalization process of each layer of the generator as an affine parameter to accomplish learning and controlling of colors. Since the depth features emphasize global semantic information, the local high-frequency information reflecting line features might be lost. A cross-layer connection structure is introduced into the generator network to strengthen the learning of line features. Experimental results show that the proposed algorithm renders line drawing of flowers better into meticulous flowers and the generated images are accordant with the color distribution and characteristics of real meticulous flowers with good artistic reality and appreciation.更多还原