Scholarlink

<p><?tk 4?>Continual learning boosts all-in-one adverse weather image restoration models’ adaptability to dynamic scenarios. However, the representative advanced method CLAIO suffers from random sampling and incomplete information preservation. To tackle these, we propose two key improvements: a representative sample selection strategy replacing random storage, which captures data distribution under fixed memory budget; and global feature constraints to build a multi-scale knowledge distillation mechanism for mitigating catastrophic forgetting. Extensive experiments show stable learning and superior performance, with 2.2% higher average PSNR. Notably, the proposed storage strategy alone brings a 10.2% PSNR gain under limited memory.</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

Continual multiple adverse weather image restoration based on multi-scale representative knowledge distillation

  • 期刊论文

  • 发表日期：

    2026年7月3日

查看全文
  • Jiayi Han,
  • Xianhua Zeng,
  • Yixin Xiang

摘要

Continual learning boosts all-in-one adverse weather image restoration models’ adaptability to dynamic scenarios. However, the representative advanced method CLAIO suffers from random sampling and incomplete information preservation. To tackle these, we propose two key improvements: a representative sample selection strategy replacing random storage, which captures data distribution under fixed memory budget; and global feature constraints to build a multi-scale knowledge distillation mechanism for mitigating catastrophic forgetting. Extensive experiments show stable learning and superior performance, with 2.2% higher average PSNR. Notably, the proposed storage strategy alone brings a 10.2% PSNR gain under limited memory.