基于注意力机制和随机像素擦除的双判别器单图生成对抗网络  

Dual Discriminator Single-image Generative Adversarial Networks Based on Attention Mechanism and Random Pixel Erasure

作　　者：彭星鸿 刘玲 袁平[1] PENG Xinghong;LIU Ling;YUAN Ping(School of Computer Science and Technology,Southwest University of Science and Technology,Mianyang 621010,Sichuan,China)

机构地区：[1]西南科技大学计算机科学与技术学院,四川绵阳621010

出　　处：《西南科技大学学报》2024年第2期100-108,共9页Journal of Southwest University of Science and Technology

摘　　要：针对单一自然图像生成模型的梯度消失和模式崩溃问题,提出了基于注意力机制和随机像素擦除的双判别器单图生成对抗网络模型。具体方法为:在生成器中引入CBAM模块以增强特征表示;采用多样性损失优化损失函数提高生成图像多样性;在图像上采样及过渡到下一阶段前增加随机像素擦除进一步丰富输出的多样性;在判别器中集成自注意力机制捕获更加全面的依赖关系;实施双判别器设计减轻模式崩溃问题。实验结果表明:与单一自然图像生成模型相比,本文方法在图像质量、多样性和训练稳定性方面均有显著提升。In addressing the issues of gradient vanishing and mode collapse in the generative model from a single natural image,we proposed the dual discriminator single-image generative adversarial networks based on attention mechanism and random pixel erasure(ADRE-SinGAN).The specific methods employed include introducing the CBAM module in the generator to enhance feature representation,optimizing the loss function with diversity loss to improve image diversity,incorporating random erasing before up-sampling and transitioning to the next stage to enrich the output diversity further,integrating self-attention mechanism in the discriminator to capture more comprehensive dependency relationships,and implementing a dual discriminator design to alleviate mode collapse issues effectively.

关 键 词：单一自然图像生成模型 注意力机制 双判别器 随机像素擦除 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]