知识卡片

普通读书笔记卡 · 8960.b · 生成模型:GPT-6

统计匹配可以学习特征

理解等级:中级

手工纹理方法预先选择滤波器及统计,GAN则通过训练判别器产生区分生成与真实样本的信号。两者都可从匹配分布特征的角度理解,但GAN并不只是精确匹配有限子带直方图,也不保证训练后已匹配所有真实图像统计。

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 Heeger-Bergen algorithm is also connected to generative adversarial networks (GANs), which we will encounter in section 32.9. The objective of Heeger-Bergen is to make a texture that has the same histogram statistics as an exemplar. A GAN also makes images that have the same statistics as a set of exemplars; however, the statistics we seek to match are learned (by a so-called discriminator neural

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第28章《纹理》(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/c028.xhtml) - 结论依据:摘录中关于统计匹配可以学习特征的定义、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 Heeger-Bergen algorithm is also connected to generative adversarial networks (GANs), which we will encounter in section 32.9. The objective of Heeger-Bergen is to make a texture that has the same histogram statistics as an exemplar. A GAN also makes images that have the same statistics as a set of exemplars; however, the statistics we seek to match are learned (by a so-called discriminator neural