基于对抗神经网络的小样本生成技术

Small-sample generation techniques based on adversarial neural networks

针对传统生成对抗网络算法使用少量训练样本无法生成高质量图像的不足,提出一种改进的数据增强对抗网络用于生成图像。改进算法利用U-Net3+结构替换原有的U-Net结构以提高模型计算效率;利用Sand Glass模块改进其残差模块以降低模型剃度混淆和信息丢失风险。为了验证所提算法的有效性,在公共数据集上进行了测试,实验结果表明,本文算法在图像生成和数据增强效果上均取得较好表现。

To propose the improved data enhancement adversarial network used to generate high-quality images using the traditional generative adversarial network algorithm.The improved algorithm uses the UNet3+ structure to replace the original U-Net structure to improve the model calculation efficiency,and improves the residual module to reduce the risk of model shaving confusion and information loss.To verify the effectiveness of the proposed algorithm,it is tested on public datasets and experimental results show that the algorithm achieves good performance in both image generation and data enhancement effects.