知识卡片
插零加卷积不是逆卷积
先在特征图样本间插零,再用可学习核滤波,可构成上采样层。它增加空间网格,常被称为UpConv或某些“反卷积”层,但并未保证反演先前卷积或恢复丢失信息。输出尺寸、相位和偏置仍需与编码路径协调。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 卷积神经网络
原书出处 · 1 处
- · 查看原书依据
Upsampling layers perform the opposite transformation, outputting a tensor that is larger in the spatial dimensions than the input: . One kind of upsampling layer can be made as the analogue of strided convolution. Strided convolution convolves then subsamples; this upsampling layer instead dilates the signal then convolves. Starting with a blank image of zeros, h = 0, we set:
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第24章《卷积神经网络》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c024b.xhtml)
- 结论依据:摘录中关于插零加卷积不是逆卷积的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:Upsampling layers perform the opposite transformation, outputting a tensor that is larger in the spatial dimensions than the input: . One kind of upsampling layer can be made as the analogue of strided convolution. Strided convolution convolves then subsamples; this upsampling layer instead dilates the signal then convolves. Starting with a blank image of zeros, h = 0, we set: