知识卡片

普通读书笔记卡 · 8912 · 生成模型:GPT-6

保留位置即可计算卷积

理解等级:中级

先卷积再抽取时,最终只需要将保留位置上的卷积结果。可直接计算这些位置的加权和以减少工作;若核可分离,还可按两个方向分解。优化要保留相同的边界、相位和归一化,不能先把原图抽稀后随意换成小网格上的同名核。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      When implementing 2D anti-aliasing filters, we apply two consecutive 1D convolutions along each dimension. For efficiency, the anti-aliasing filter only needs to be computed on the samples that will be kept after downsampling (k-strided convolution). The necessary amount of attenuation of the aliasing frequencies will have to be decided based on each application.

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第21章《图像下采样与上采样》(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/c021.xhtml) - 结论依据:摘录中关于保留位置即可计算卷积的定义、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:When implementing 2D anti-aliasing filters, we apply two consecutive 1D convolutions along each dimension. For efficiency, the anti-aliasing filter only needs to be computed on the samples that will be kept after downsampling (k-strided convolution). The necessary amount of attenuation of the aliasing frequencies will have to be decided based on each application.