← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8909 · 生成模型:GPT-6

步长卷积仍需抗混叠

理解等级:中级

步长为 k 的卷积可拆成先卷积、再每隔 k 个位置抽取输出。卷积本身不自动保证抗混叠,尤其导数核还可能增强高频。分析网络中的分辨率变化,应检查抽取之前的频率限制,而不能因为它是可学习层就假定保留了正确纹理。

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

      This system has three steps: the first step is a vertical edge filter (convolution of the input, ℓin, with a horizontal kernel, [1, –1]). The next step is to reduce the resolution of the filter output by taking one pixel every two along each dimension (this operation is called decimation). Decimation takes an input, u, with size N × M, and outputs, h, with half

    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:摘录中关于步长卷积仍需抗混叠的定义、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:This system has three steps: the first step is a vertical edge filter (convolution of the input, ℓin, with a horizontal kernel, [1, –1]). The next step is to reduce the resolution of the filter output by taking one pixel every two along each dimension (this operation is called decimation). Decimation takes an input, u, with size N × M, and outputs, h, with half