基于注意力机制的生成对抗网络古诗插图生成方法  

Generation Method of Ancient Poetry Illustrations Based on Attention Mechanism of Generative Adversarial Network

作　　者：沈协伟 魏为民 SHEN Xiewei;WEI Weimin(School of Computer Science and Technology,Shanghai University of Electric Power,Shanghai 200090,China)

机构地区：[1]上海电力大学计算机科学与技术学院,上海200090

出　　处：《上海电力大学学报》2023年第2期175-181,共7页Journal of Shanghai University of Electric Power

摘　　要：古诗插图自动生成是非常具有挑战性的任务。提出了一种新的方法,通过使用注意力机制的生成对抗网络,输入古诗文字和草绘图,经由网络输出古诗插图。该模型在利用编码与解码的图片生成结构的同时,结合注意力机制,将图像色彩及轮廓特征进行了提取,生成器中加入了残差网络模块,增加了网络深度,同时提高了网络模型的效果,最终生成图像质量更高的古诗插图。实验结果表明,所提方法相较于对比方法,在图像质量和色彩渲染方面均有优势,输入的古诗和草图经过网络模型,生成了较为符合古诗意境的插图。It is a challenging task to realize the automatic generation of ancient poetry illustrations.This paper proposes a generation method using attention mechanism against network generation,which outputs ancient poetry illustrations through the network by inputting ancient poetry text and sketching.The model uses the image generation structure of encoding and decoding,and extracts the image color and contour features combined with the attention mechanism.The residual network module is added to the generator,which can increase the network depth,improve the effect of the network model,and finally generate ancient poetry illustrations with higher image quality.The experimental results show that the proposed method has advantages in image quality and color rendering compared with the comparison method.The input ancient poems and sketches can generate illustrations that are more consistent with the artistic conception of ancient poems through the network model.

关 键 词：古诗插图 图像生成 注意力机制 生成对抗网络 残差网络模块 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]