主要的亮点在于各个encoder和decoder中的用于取代self-attention的local MLP和global MLP的Multi-axis gated MLP block (MAB)。而这个MAB的block,思想其实也跟Swin-transformer的窗口化操作很类似,只是本文是在两个不同轴上进行,一个是在窗口内部,引入local interaction;另一个是在窗口之间,引入global interaction。这种...
D. 使用MAXIM-1S 和 MAXIM-2S 在SIDD 去加入MAXIM-FFT, -MLP, -gMLP, -SA。加入gMLP和SA效果比较好。 3.2Comparison with Other MLPs 我们对最近的MLP模型中有效感受野进行了可视化比较。 3.3 去噪在SIDD, DND 数据集 3.4 去模糊在GoPro, HIDE,RealBlur数据集 3.5 去雨在Rain100L, Rain100H, Test100, ...
Multi-Axis Gated MLP(MAB) 这份工作收到了Improved Transformer for High-Resolution GANs(HiT)中引入的多轴块自注意力机制的启发,但是这篇参考工作中的设计并不太适合与图像恢复或者是图像增强任务,因为要面对的图像通常具有任意的形状(这里解释的有些牵强,因为严格来讲,本文的方法也没法处理任意形状,毕竟每个模块的...
MAXIM: Multi-Axis MLP for Image Processing 9 January, 2022 CVPR 2022 Oral (in 33 best paper finalist) https://arxiv.org/abs/2201.02973 https://github.com/google-research/maxim Authors: Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, Yinxiao Li ...
提出另一个「即插即用」的交叉门控模块(Cross-Gating MLP block),可以无痛替代交叉注意力机制,并且同样在线性复杂度享有全局 / 局部感受野和全卷积特性。 MAXIM: Multi-Axis MLP for Image Processing 论文地址:arxiv.org/abs/2201.02973 代码/模型/实验结果:...
在CVPR 2022年的 Oral 会议上，MAXIM：Multi-Axis MLP for Image Processing引起了关注，它展示了一种创新的底层视觉任务处理架构，刷新了多项SOTA记录。MAXIM基于U-net结构，采用了一个多轴门控MLP和一个交叉门控块，旨在结合局部和全局视觉信息，同时保持低复杂度和全卷积特性，适用于去噪、去模糊等低...
In this work, we present a multi-axis MLP based architecture called MAXIM, that can serve as an efficient and flexible general-purpose vision backbone for image processing tasks. MAXIM uses a UNet-shaped hierarchical structure and supports long-range interactions enabled by spatially-gated MLPs. ...
In this work we present a multi-axis MLP based architecture, called MAXIM, that can serve as an efficient and flexible general-purpose vision backbone for image processing tasks. MAXIM uses a UNet-shaped hierarchical structure and supports long-range interactions enabled by spatially-gated MLPs. ...