文章目录
1. Introduction
What is Digital Image Processing?
\quad Improvement of pictorial information for human interpretation.
\quad Processing of image data for storage, transmission and representation for autonomous machine perception.
\quad 为了人的视觉感官服务或者为计算机存储、传输、表达图像而进行预处理。
Key Stages in Digital Image Processing
步骤依次如下:
- Image Acquisition(获取图像)
\quad 主要是完成成像的工作。 - Image Enhancement(图像增强)
\quad 比如提高对比度。 - Image Restoration(图像复原)
\quad 比如去噪。 - Morphological Processing(形态学操作)
- Segmentation(图像分割)
\quad 提取图像感兴趣的部分。 - Representation & Description(图像描述)
\quad 提取图像感兴趣的特征。 - Object Recognition(目标识别)
2. Fundamentals
- Digitizing the coordinate values is called sampling.(采样)
Digitizing the amplitude is called quantization.(量化) - DPI(dots per inch)
\quad 度量分辨率用的。 - Image Interpolation
\quad nearest neighbor, bilinear and bicubic
\quad 插值算子描述(引用) - Image Resolution(图像分辨率)
\quad The number of bits used to quantize intensity is often referred as the intensity resolution.
\quad 例子:强度等级有256种,强度分辨率为8。
3. Intensity Transformations
About image enhancement
\quad Highlighting interesting details in images.
\quad Removing noise from images.
\quad Making images more visually appealing.
\quad 高亮感兴趣的细节、去除噪声、让图像可视性更强。
Some basic intensity transformation functions
- Image Negatives
s = T ( r ) = L − 1 − r s= T(r) =L-1-r s=T(r)=L−1−r - Thresholding
s = { 1 r > t h r e s h o l d 0 r > t h r e s h o l d s= \begin{cases} 1 & r>threshold \\ 0 &r>threshold\end{cases} s={ 10r>thresholdr>threshold
人们更倾向于在白色的背景下,分析黑色物体的细节。
- Log Transformations
s = c log v + 1 ( 1 + v r ) s=c \log_{v+1}{(1+vr)} s=clogv+1(1+vr) - Gamma Transformations
s = c r γ s=cr^{\gamma} s=crγ
3和4中的 r r r倾向于范围在 [ 0 , 1 ] [0,1] [0,1]之间
Histogram Equalization
\quad 核心: p s ( s ) = 1 p_s(s)=1 ps(s)=1
注意为概率密度函数,不是概率累积函数,此处s的取值范围为[0,1]。
我们希望等概率出现,所以概率密度函数为1。若深究概率累积函数,则为:
P s ( s ) = s 0 ≤ s ≤ 1 P_s(s)=s \quad \quad 0 \leq s \leq 1 Ps(s)=s0≤s≤1
- The discrete version of the cumulative distribution function
P s ( s ) = ∑ i = 0 k P k ( k ) P_s(s)=\sum_{i=0}^{k}P_k(k) Ps(s)=∑i=0kPk(k) - The discrete version of the transformation function
s k = I N T [ ( L − 1 ) T ( k ) ] s_k=INT[(L-1)T(k)] sk=INT[(L−1)T(k)]
w h e r e T ( k ) = ∑ i = 0 k P ( i ) = ∑ i = 0 k n i n where \quad T(k)=\sum_{i=0}^{k}P(i)=\sum_{i=0}^{k}\dfrac{n_i}{n} whereT(k)=∑i=0kP(i)=∑i=0knni
4. Spatial Filtering
Smoothing Spatial Filters
\quad The output of a linear smoothing spatial filter is simply the (weighted) average of the pixels in the neighborhood defined by the filter mask; sharp transitions are reduced.
\quad Linear smoothing filters are low‐pass filters; they will blur edges.
\quad They can reduce irrelevant details in an image and get a gross representation of objects of interest.
\quad 降低噪声(及细节)、模糊边界(较大的梯度变化会受到抑制)、属于低通滤波器。
- Types of image noise.
\quad Salt and Pepper & White,Gaussian
\quad 椒盐噪声、高斯白噪声 - Filter
\quad Mean Filter:
\quad [ 1 1 1 1 1 1 1 1 1 ] \begin{bmatrix} 1 & 1 & 1 \\ 1 & 1 & 1 \\ 1 & 1 & 1 \end{bmatrix} ⎣⎡111111111⎦⎤
\quad Improved Filter:
\quad [ 1 2 1 2 4 2 1 2 1 ] \begin{bmatrix} 1 & 2 & 1 \\ 2 & 4 & 2 \\ 1 & 2 & 1 \end{bmatrix} ⎣⎡121242121⎦⎤
\quad Gaussian Filter:
\quad f ( x , y ) = 1 2 π σ e − ( x − x 0 ) 2 + ( y − y 0 ) 2 2 σ 2 f(x,y)=\cfrac{1}{2\pi\sigma}e^{-\cfrac{(x-x_0)^2 +(y-y_0)^2}{2\sigma^2}} f(x,y)=2πσ1e−2σ2(x−x0)