基于生成对抗网络的自适应归一化数据增强算法
Adaptive normalized data augmentation algorithm based on DCGAN
赵彬 1高永乐 1王清璇 1王泽 1张钧溟1
作者信息
- 1. 长春工业大学电气与电子工程学院,吉林长春 130012
- 折叠
摘要
提出一种基于深度卷积生成对抗网络的自适应归一化数据增强算法,通过深度卷积神经网络强大的特征提取能力提升了生成图像的真实度,进一步采用自适应实例归一化解决因梯度爆炸导致的网络过拟合问题,最后在CIFAR10数据集上进行模型训练,并对输入的原始样本图像进行1 000轮对抗得到生成图片,实验结果显示,文中提出的自适应归一化方法生成图像真实度提升1.2％.
Abstract
To address this issue,this paper proposes an adaptive normalized data augmentation algorithm based on DCGAN.Firstly,the powerful feature extraction capability of the DCGAN improves the realism of the generated images.Subsequently,using adaptive instance normalization to solve the problem of network overfitting caused by gradient explosion.On this basic,the model was trained on the CIFAR10 dataset,and the input original sample image was subjected to 1 000 rounds of confrontation to obtain the generated image.Finally,experimental results show that the adaptive normalization method proposed in this article improves the authenticity of the generated images by 1.2％.
关键词
自适应归一化/数据增强/深度卷积生成对抗网络/深度学习Key words
adaptive normalization/data augmentation/DCGAN(Deep Convolutional Generative Adversarial Network)/deep learning引用本文复制引用
出版年
2024
NETL
NSTL
维普
万方数据