Install Cpp jupyter kernel, Image Guied Filter

Install Cpp kernel

• jupyter notebook 安装 C/C++ kernel - KEN的文章 - 知乎
• conda install xeus-cling -c conda-forge
• after install
  • jupyter kernelspec list
  • fix issue

  • by

    cp /opt/anaconda3/envs/gpu/lib/python3.8/_sysconfigdata_x86_64_conda_cos6_linux_gnu.py /opt/anaconda3/envs/gpu/lib/python3.8/_sysconfigdata_x86_64_conda_linux_gnu.py
    

  • check installation

    python3    /opt/anaconda3/envs/gpu/share/jupyter/kernels/python3
    xcpp11     /opt/anaconda3/envs/gpu/share/jupyter/kernels/xcpp11
    xcpp14     /opt/anaconda3/envs/gpu/share/jupyter/kernels/xcpp14
    xcpp17     /opt/anaconda3/envs/gpu/share/jupyter/kernels/xcpp17
    

Guided Image Filtering

• 
  • PPT eccv 2010
  • Paper.pdf - Microsoft Research

Image Filtering

• Explicit LTI [linear translation-invariant, Linear time-invariant]
• Explicit Weighted-Average Filters
  • Gaussian filter
  • Laplacian filter
  • Sobel filter
• the process of solving a Poisson Equation
  • HDR compression[high dynamic range][4]
  • image stitching[5]
  • image matting[6]

Related

• Explicit Weighted-Average Filters
  • Bilateral Filter
    • 距离参数*差异参数，作为权重

    • The bilateral filter computes the filter output at a pixel as a weighted average of neighboring pixels

    • 看起来像是二阶过滤？
    • time cost
      • brute-force, $O(Nr^2)$
      • [15,16], $O(N)$
      • [17]Adams et al. color image
      • $r$: kernel radius

      • All the abovemethods require a high quantization degree to achieve satisfactory speed, but at the expense of quality degradation

    • problem
      • gradient reversal
      • 
• Optimization-based Image Filtering

  • generate high quality results, solving the corresponding linear system is time-consuming.

• Nonaverage Filters, 非均质滤波, 非线性滤波
  • 中值滤波
  • Total-Variation filter
• Guided Filter
  • 
  • 
  • 
  • term
    • $q$: output Image
    • $p$: input Image
    • $I$: guide Image
    • $w$: window
    • $k$: the k_th windows
    • $a,b$: linear model params
    • $n$: noise or texture information
  • principle
    • $$ \begin{aligned} q_i &= a_kI_i + b_k, \forall i \in w_k&(1)\\ q_i & = p_i - n_i&(2)\\ E(a_k,b_k) &=\sum_{i\in w_k}\big((a_k I_i + b_k - p_i)^2\big)+\epsilon a_k^2&(3)\\ a_k&=\frac{\frac{1}{\vert w\vert}\sum_{i \in w_k }I_i p_i - \mu_k\bar{p_k}}{\sigma_k^2+\epsilon}&(4)\\ b_k&=\bar{p_k}-a_k\mu_k&(5)\\ q_i&=\frac{1}{\vert w\vert}\sum_{i \in w_k }a_k I_i+b_k&(6)\\ \end{aligned} $$
    • 其中
      • $(1)$: 期望的引导图和输入图的线性关系
      • $(2)$: 输入和输出之间差一个噪声或纹理
      • $(3)$: 领回归计算$a,b$的值
      • $(4,5,6)$: 参数和输出图片计算
  • evaluation
    • PSNR [Peak signal-to-noise ratio]
      • $$ \begin{aligned} PSNR&=10\log_{10}\frac{MAX_I^2}{MSE}\\ &=20\log_{10}\frac{MAX_I}{\sqrt{MSE}}\\ &=20\log_{10}MAX_I-10\log_{10}{MSE}\\ \end{aligned} $$

Code

• mylikes
  • 
• cat guide filter
  • 4
• cat bi filter
  • 
• 

def guider_filter_image(image_name:str):
    img = cv2.imread(image_name)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    radius = [2**x for x in range(5)]
    epsilon = [.1**x for x in range(5)]
    row = len(radius)
    col = len(epsilon)
    guided = []
    titles = []
    for p in tqdm(cartesian_product([radius, epsilon])):
        r, e = p
        title = f"r:{r}_e:{e:.0e}"
        guided.append(guidedFilter(guide=img, src=img, radius=int(r), eps=e))
        titles.append(title)

    fig = px.imshow(
        np.array(guided),  
        facet_col_wrap = col,
        facet_col = 0,
    )

    for i, label in enumerate(titles[::-1]):
        fig.layout.annotations[i]['text'] = label

    fig.update_xaxes(showticklabels=False).update_yaxes(showticklabels=False)
    fig.update_layout(
        autosize=False,
        width=720,
        height=720,
        margin=dict(l=0,r=0,b=0,t=0, pad=0,autoexpand=False))
    fig.show()

Reference

• 导向滤波原理（Guided Filter）- CSDN - [江户川柯壮]
• 导向滤波(Guided Filter)公式详解 - CSDN - [lsflll]