Scholarlink

<p>In recent years, knowledge distillation methods applied to Generative Adversarial Networks (GANs) have faced deployment challenges due to the limitations of single-network feature extraction, which hinders the capture of multi-scale hierarchical features and the establishment of cross-semantic dependencies. This results in degraded knowledge transfer quality from teacher to student models, poor generalization capability, and training instability. To address these issues, we propose a Dynamic Fusion of Multi-Scale Perception and Adaptive Discrimination(DFAD) method. DFAD employs dynamic weighting to reconcile architectural differences and minimize the perceptual gap between teacher and student models, thereby enhancing feature decoupling. Furthermore, an adaptive convolutional module with a channel attention mechanism dynamically adjusts the importance of feature maps, improves discriminator flexibility, and alleviates mode collapse. Experiments demonstrate that DFAD reduces CycleGAN’s computational cost to 1/40th (measured in multiply-accumulate operations) and parameter count to 1/80th, while achieving an FID score of 72.91 (surpassing the current SoTA 73.54). DFAD also maintains competitive performance on Pix2Pix.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dynamic fusion of multi-scale perception and adaptive discrimination for compressed GANs

  • 期刊论文

  • 发表日期：

    2025年7月2日

查看全文
  • Rui Li,
  • ChangHao Ge

摘要

In recent years, knowledge distillation methods applied to Generative Adversarial Networks (GANs) have faced deployment challenges due to the limitations of single-network feature extraction, which hinders the capture of multi-scale hierarchical features and the establishment of cross-semantic dependencies. This results in degraded knowledge transfer quality from teacher to student models, poor generalization capability, and training instability. To address these issues, we propose a Dynamic Fusion of Multi-Scale Perception and Adaptive Discrimination(DFAD) method. DFAD employs dynamic weighting to reconcile architectural differences and minimize the perceptual gap between teacher and student models, thereby enhancing feature decoupling. Furthermore, an adaptive convolutional module with a channel attention mechanism dynamically adjusts the importance of feature maps, improves discriminator flexibility, and alleviates mode collapse. Experiments demonstrate that DFAD reduces CycleGAN’s computational cost to 1/40th (measured in multiply-accumulate operations) and parameter count to 1/80th, while achieving an FID score of 72.91 (surpassing the current SoTA 73.54). DFAD also maintains competitive performance on Pix2Pix.