Kivonat
最近、CNNとトランスフォーマーに基づく画像復元の重要な進展がありました。しかし、多くの既存の研究は、IRタスクの特性を考慮せずに基本的なブロックの設計に焦点を当てています。このアプローチはパラメータ冗長性と不必要な計算をもたらし、画像復元の効率が妨げられます。そこで、著者らはLIRと呼ばれる軽量な画像復元ベースラインを提案しました。LIRは、局所およびグローバル残差接続に存在する劣化を取り除き、Adaptive FiltersとAttention Blocksから構成されるLightweight Adaptive Attention(LAA)ブロックを導入しています。これにより、LIRは一部のタスクで最先端モデルと比較可能な性能を達成しました。
Statisztikák
LIR achieves comparable performance to state-of-the-art models with fewer parameters and computations in certain tasks.
LIR produces better visual results than state-of-the-art networks that are more in line with the human aesthetic.
LIR achieves state-of-the-art on SSIM, significantly outperforming Restormer.
LIR significantly outperforms SFNet with fewer parameters and computations.
LIR achieves comparable performance to the state-of-the-art method on PSNR and the best performance on SSIM with smaller parameter numbers and computations.
Idézetek
"Many existing works only focus on designing the basic block and inefficiently stacking these blocks in a model, without considering the specific characteristics of the IR task."
"Our goal is an efficient and lightweight network design for various IR tasks."
"LIR addresses the degradations existing in the local and global residual connections that are ignored by modern networks."
"LAA is capable of adaptively sharpening contours, removing degradation, and capturing global information in various Image Restoration scenes in a computation-friendly manner."
"LIR achieves comparable performance to state-of-the-art models in PSNR and produces better visual results that are more in line with human aesthetic."