基于生成式对抗网络的复杂结构图像生成方法  

Generative Adversarial Network-Based Method for Generating Complex Structured Images

作　　者：陈颀周 邵清[1] 卢军国 李全全 CHEN Qizhou;SHAO Qing;LU Junguo;LI Quanquan(School of Optical-Electrical&Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093;Shanghai Waigaoqiao Shipbuilding Co.,Ltd.,Shanghai 200137)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093 [2]上海外高桥造船有限公司,上海200137

出　　处：《计算机与数字工程》2024年第8期2457-2463,2492,共8页Computer & Digital Engineering

基　　金：国家自然科学基金项目(编号:61502220);上海市科学教育委员会(编号:19511105103)资助。

摘　　要：在图像生成任务中,生成式对抗网络通常更倾向于学习图像的纹理特征等低级模式,而忽略其形状特征等高级模式,导致生成不规则图像。为了解决这一问题,论文提出残差注意力多通道生成式对抗网络。首先,利用残差连接,使模型信息的传输效率变得更高,模型训练变得更加稳定。第二,引入自注意力机制加强生成图像的长距离依赖,提升模型对形状的建模能力。第三,利用HS通道隐藏亮度细节,而突出图中对象轮廓的特点,将HS通道输入判别器,作为额外判别依据,使模型的FID分数获得了进一步提升。实验结果表明,在几何结构复杂的数据集上,该模型能够生成视觉效果更好、更符合现实情况的图像。In image generation tasks,generative adversarial networks usually prefer to learn low-level patterns,such as textures of images,but ignore high-level patterns,such as their shapes.To solve this problem,this paper proposes a residual attention multi-channel generative adversarial network.Firstly,the residual connection is used to make information transmission more efficient and the model training becomes more stable.Secondly,the self-attention mechanism is introduced to improve the long-range dependence of the generated images and enhance the modeling capacity of the model on shapes.Thirdly,the model's FID score is further improved by using the concatenation of RGB and HS channels of an image as the input of the discriminator,where the HS channels hides the luminance details while highlighting the object contours in the figure.The experimental results show that the model is able to generate images with better visual effects on datasets with complex geometric structures.

关 键 词：图像生成 生成式对抗网络 残差网络 自注意力机制 颜色通道 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]