← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8914.c · 生成模型:GPT-6

重采样伪影会贯穿网络

理解等级:中级

视觉网络中的特征图缩放、图像金字塔以及生成模型的逐级放大都会受到采样与滤波选择影响。别名可产生伪纹理或降低对平移的稳定性。逐阶段规定滤波、网格和边界能减少这些问题,但抗混叠并不自动保证完全平移不变或消除所有生成伪影。

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

      Aliasing is a constant source of problems in many image processing pipelines. In computer vision, aliasing can introduce noise, spurious features and textures, and lack of translation invariance. In chapter 24 we will discuss how aliasing affects the quality of some of the convolutional layers used in neural networks.

    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:摘录中关于重采样伪影会贯穿网络的定义、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:Aliasing is a constant source of problems in many image processing pipelines. In computer vision, aliasing can introduce noise, spurious features and textures, and lack of translation invariance. In chapter 24 we will discuss how aliasing affects the quality of some of the convolutional layers used in neural networks.