← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8876.a · 生成模型:GPT-6

卷积滑动内积前需反转核

理解等级:中级

卷积可看作反转核再滑动做内积。一维h[−1]=1、h[0]=2、h[1]=3时,输出为3ℓ[n−1]+2ℓ[n]+ℓ[n+1]。二维同时反转两条空间轴。对非对称核漏掉反转会得到不同运算,方向导数的符号也可能改变。

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

      The convolution operation can be described in words as: first take the kernel h [k] and mirror it h [–k], then shift the mirrored kernel so that the origin is at location n, then multiply the input values around location n by the mirrored kernel and sum the result. Store the result in the output vector ℓout [n]. For instance, if the convolutional kernel has

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第15章《线性图像滤波》(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/c015b.xhtml) - 结论依据:摘录中关于卷积滑动内积前需反转核的定义、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:The convolution operation can be described in words as: first take the kernel h [k] and mirror it h [–k], then shift the mirrored kernel so that the origin is at location n, then multiply the input values around location n by the mirrored kernel and sum the result. Store the result in the output vector ℓout [n]. For instance, if the convolutional kernel has