基于CartoonGan的改进卡通化图片生成方法  

Improved Cartoon Image Generation Method based on GartoonGan

作　　者：张文天 于瓅[1] ZHANG Wen-tian;YU Li(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan,Anhui,232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《新疆师范大学学报（自然科学版）》2024年第2期32-42,共11页Journal of Xinjiang Normal University(Natural Sciences Edition)

基　　金：2021年安徽省重点研究与开发计划项目(202104d07020010)。

摘　　要：文章提出了一种改进的卡通图像生成网络模型,旨在增强卡通效果的同时保留语义信息。首先,设计一个显著性网络,为了防止过拟合问题以及进一步提取特征,在显著性网络上添加残差结构构成残差显著性网络,并将其拼接到CartoonGan上用来保留语义信息;其次,在前者基础上添加cbam注意力机制进一步提高卡通化效果;最后,在训练过程中为了防止训练不稳定和梯度消失使用最小二乘损失来替换交叉熵损失,并且引入显著性损失来约束显著性网路的训练。实验表明,通过在宫崎骏和新海诚两个画风的卡通数据集上进行测试,使用FID测试指标显示在两个数据集上都有一定的优化。This article proposes an improved cartoon image generation network model aimed at enhancing cartoon effects while preserving semantic information as much as possible.Firstly,a saliency network was designed and in order to prevent overfitting problems and extract further features,residual structures were added to the saliency network to form a residual saliency network and concatenated onto CartoonGan to preserve semantic information;Secondly,adding a cbam attention mechanism to the former further improves the cartoonization effect;Finally,in order to prevent training instability and gradient vanishing during the training process,the least squares loss is used to replace the cross entropy loss,and the significance loss is introduced to constrain the training of the significance network.The experiment showed that through testing on cartoon datasets of Miyazaki Hayao and Makoto Shinkai,the use of FID testing indicators showed some optimization on both datasets.

关 键 词：Cbam注意力机制 显著性网络 残差结构 CartoonGan 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]