Delving Deeper into Anti-Aliasing in ConvNets
Delving Deeper into Anti-Aliasing in ConvNets
复制标题
DOI：
10.1007/s11263-022-01672-y
复制
发表时间：
2020-08
影响因子：
19.5
通讯作者：
Xueyan Zou;Fanyi Xiao;Zhiding Yu;Yong Jae Lee
中科院分区：
文献类型：
--
作者：
Xueyan Zou;Fanyi Xiao;Zhiding Yu;Yong Jae Lee
混叠是指高频信号在采样后退化为完全不同的信号的现象。这是深度学习背景下的一个问题，因为深度架构中广泛采用下采样层来减少参数和计算。标准解决方案是在下采样之前应用低通滤波器（例如高斯模糊）（Zhang in：ICML，2020）。然而，在整个内容上应用相同的过滤器可能不是最理想的，因为特征图的频率可能会随着空间位置和特征通道的不同而变化。为了解决这个问题，我们提出了一个自适应内容感知低通滤波层，它为输入特征图的每个空间位置和通道组预测单独的滤波器权重。我们研究了所提出的方法在多个任务中的有效性和泛化性，包括图像分类、语义分割、实例分割、视频实例分割和图像到图像翻译。定性和定量结果都表明，我们的方法有效地适应不同的特征频率，以避免混叠，同时保留有用的识别信息。代码可在 https://maureenzou.github.io/ddac/ 获取。
Aliasing refers to the phenomenon that high frequency signals degenerate into completely different ones after sampling. It arises as a problem in the context of deep learning as downsampling layers are widely adopted in deep architectures to reduce parameters and computation. The standard solution is to apply a low-pass filter (e.g., Gaussian blur) before downsampling (Zhang in: ICML, 2020). However, it can be suboptimal to apply the same filter across the entire content, as the frequency of feature maps can vary across both spatial locations and feature channels. To tackle this, we propose an adaptive content-aware low-pass filtering layer, whichpredicts separate filter weights for each spatial location and channel groupof the input feature maps. We investigate the effectiveness and generalization of the proposed method across multiple tasks, including image classification, semantic segmentation, instance segmentation, video instance segmentation, and image-to-image translation. Both qualitative and quantitative results demonstrate that our approach effectively adapts to the different feature frequencies to avoid aliasing while preserving useful information for recognition. Code is available at https://maureenzou.github.io/ddac/.