This paper outlines:
A computer vision (hereinafter referred to as CV) Overview
Second, image preprocessing
o image display and storage principles (slightly)
o image enhancement goals
The image processing method o
o point operation: based on a histogram of contrast enhancement
o morphology process (abbreviated)
o spatial domain processing: convolution
o Convolution application (smoothing, edge detection, sharpening, etc.)
o Frequency domain processing: Fourier Fourier transform, wavelet transform (slightly)
O applications:
smoothing, edge detection, etc. CLAHE
A, CV research content:
The next update will be included in the content of the article:
Second, image preprocessing
o Image enhancement goals
The image processing method o
Because CNN or other deep learning networks, are required for picture data for feature extraction in order to learn. Therefore extracted picture feature is very important, feature extraction following a brief introduction:
o point operation: histogram-based contrast enhancement
Explanation:
A statistical picture data / distribution characteristics, personal understanding a little cluster of ideas, just like a red and blue T-shirt, histogram statistics, finally can get T-shirts (here only by color, for example) histogram color distribution. This is a way to extract its features.
Here, " gray ": the midpoint of the black and white image color depth, 0 to 255, 0 is black, 255 is white. Gray scale digital image is an image of only one sample per pixel color.
Each pixel : 256 gray levels (pixel values between 0 and 255), each representing a different level of brightness.
Explanation:
The larger (maximum 255) taken bin, wherein the distribution is dispersed in advance, for each pixel figured out, corresponds to only one section (assuming take 255),
The smaller the bin taken, such as to take 10 in the section 10 which may comprise a plurality of different values of pixels.
Why should the histogram equalization:
这是为了对图片进行调整，所拍的图，往往因为太亮或者太暗等因素造成了，局部的灰度集中化，因此需要一种手段，进行（非线性的）重新分配各个灰度单位中的像素点数量（具体方法，一般是最终使得一定灰度范围内像素点数量的值大致相同）
AHE的问题是：经过均衡化的图像往往不连续
o 空间域处理：卷积
o 卷积的应用（平滑、边缘检测、锐化等）
因为均值滤波有很大的缺陷，因此用的很少，下面介绍几种常见的滤波器
中值滤波：有效去除椒盐噪声
高斯滤波（重点）：高斯滤波名字的由来，我的理解是，它的滤波器（卷积核）数值分布是高斯分布，它的应用必须结合后面要讲的高斯金字塔（捕捉不同尺寸的物体，尺寸缩小 + 高斯模糊）一起讲。
这里简单说明，高斯滤波其实是高斯模糊（blur），对于设定不同的标准差，高斯滤波将图片变换成不同清晰度。
标准差越大，大部分数值与均值差异较大，数据分布越离散，所用的高斯滤波器（卷积核）分布越尖锐（想象高斯分布尖尖的突出的曲线），所得的图片就会越模糊（但是所得的图像依旧是平缓的，是大致可以看出是什么物体的）。标准差越小则相反，高斯滤波器分布越平缓，所得图片不会那么模糊。
还有一些做边缘检测，团块检测，Laplacian 滤波锐化等滤波器，这里不做一一介绍，详情参考PPT
直接进入重点：高斯金字塔
高斯金字塔：通过高斯平缓(模糊) + 降采样(尺寸缩小)， 达到捕捉不同尺寸物体的目的。但是如果还想回复原图像怎么办，这就需要拉普拉斯金字塔啦。
拉普拉斯金字塔：将小尺寸的图像数据，进行上采样，得到大尺寸的图像。具体看下图中的说明：
