| 摘 要: 针对轻量化生成对抗网络FastGAN在小样本高分辨率图像生成任务中存在的深层特征语义信息丢失、细节生成能力不足、跨尺度特征融合不充分的问题,提出多维自适应特征增强模块,通过多尺度记忆引导融合机制实现跨层级特征有效传递,结合方差感知的通道-空间双维度自适应加权实现关键特征动态增强,最后引入跨尺度方差驱动的特征校准分支优化特征分布。在9组覆盖256×256分辨率与1024×1024分辨率的小样本数据集上开展实验,结果表明,改进后算法的FID指标较原FastGAN平均降低6.3%,单数据集最大降幅达21.8%,生成图像的细节清晰度与语义一致性显著提升,可为小样本高分辨率图像生成任务提供技术支撑。 |
| 关键词: 图像生成 生成对抗网络 特征增强 小样本学习 深度学习 |
|
中图分类号: TP391
文献标识码:
|
|
| Improved FastGAN High-Resolution Few-Shot Image Generation Algorithm Based on Multi-dimensional Adaptive Feature Enhancement |
|
CHEN Borun
|
School of Computer Science and Communication Engineering, Jiangsu University
|
| Abstract: Aiming at the problems of deep feature semantic information loss, insufficient detail generation ability and inadequate cross-scale feature fusion of the lightweight generative adversarial network FastGAN in high-resolution few-shot image generation tasks,a multi-dimensional adaptive feature enhancement module is proposed.The module realizes effective transmission of cross-level features through multi-scale memory-guided fusion mechanism,realizes dynamic enhancement of key features combined with variance-aware channel-spatial dual-dimensional adaptive weighting,and optimizes feature distribution by introducing cross-scale variance-driven feature calibration branch.Experiments are carried out on 9 groups of few-shot datasets covering 256×256 and 1024×1024 resolutions.The results show that the average FID index of the improved algorithm is reduced by 6.3% compared with the original FastGAN,and the maximum reduction of a single dataset reaches 21.8%.The detail clarity and semantic consistency of the generated images are significantly improved,which can provide technical support for high-resolution few-shot image generation tasks. |
| Keywords: image generation generative adversarial network feature enhancement few-shot learning deep learning |
