知识卡片

普通读书笔记卡 · 8954.c · 生成模型:GPT-6

拉普拉斯先验产生软阈值

理解等级:高级

观测系数y=x+ε且ε方差为τ²,先验p(x)正比exp(−|x|/s)时,MAP最小化(y−x)²/(2τ²)+|x|/s,得到sign(y)max(|y|−τ²/s,0)。小响应归零,大响应向零收缩;阈值来自该似然和先验尺度,不能任意沿用像素噪声方差。

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 highly kurtotic distribution of band-pass filter responses in images provides a regularity that is very useful for image denoising. If we assume zero-mean additive Gaussian noise of variance σ2, and the observed subband coefficient value , where is the subband of the observed noisy input image, then the likelihood of any coefficient value, x = ℓ0[n, m], is

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第27章《统计图像模型》(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/c027b.xhtml) - 结论依据:摘录中关于拉普拉斯先验产生软阈值的定义、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 highly kurtotic distribution of band-pass filter responses in images provides a regularity that is very useful for image denoising. If we assume zero-mean additive Gaussian noise of variance σ2, and the observed subband coefficient value , where is the subband of the observed noisy input image, then the likelihood of any coefficient value, x = ℓ0[n, m], is