[Go through 3] Convolution & Image Noise & Edge & Texture

早上

Get nine hours of sleep
An eye-opening thing,Somehow clickedqqGo in the group,薅羊毛,Shopping for an hour,Then it's cooking（早午饭）,吃了之后,I explained it to my girlfriend againleetcodeThe point mechanism and myself are too bad,So don't do that.
Then a brush The problem of finding the sum of subtrees,后序遍历,然后记录,去重.
Next is to look at the courseware of deep learning

看课件

卷积核

定义上,Convolution requires the convolution kernel first 翻转180° 体现在-u -v ,如果是 +u +v It is correlation not convolution.
但是 Generally, the convolution kernel is symmetric,所以无所谓了,然后 What the neural network learns is also reversed,So no emphasis

Pan after rolling,And convolve after translation,一样的.然后 into the convolution kernel,Just make sure that9A grid of data will not be destroyed

inherent problem–边界填充

通常都是0填充 This ensures that the input and output patterns are the same size
当然也可以 镜像填充,拉伸填充

通过（消除）noise to elicit convolution（的概念）
Then there is boundary padding to elicit different convolution kernels

卷积核的尺寸 is an inherent parameter 33 55 7*7

求平均-Get closer to yourself and your neighbors,make it similar,平滑

Different convolution kernels can achieve different image operations

原图+边缘图=锐化

平均卷积核 肯定是不太行的,Bring out the Gaussian convolution kernel

中心是(0,0)
to be unified,不然 The original image may be enlarged or reduced 数据

方差越小,The steeper the peak,越集中,The more inhumane（far from neighbors）,The smoothing effect is less obvious

模板尺寸,对应着 Normalized denominator.

经验之谈

And this convolution kernel can be separated


Multiple convolutions with a small template are less computationally intensive than a single convolution with a large template
This can be further reduced with decomposition templates

Therefore, noise can be suppressed,实现平滑（and can be separated to reduce complexity）

噪音

高斯噪音 密密麻麻 随机分布,It is equivalent to superimposing a randomly distributed signal of Gaussian distribution on the original signal

A large convolution kernel is used to remove Gaussian noise（Going to get a peel）

Gaussian convolution kernels are linear operations,But the median filter just sorts and takes the median,所以他不是

边缘

目的

边缘 A point where the brightness changes sharply

Different application scenarios focus on different edges

• Determine what the object is,You need to recognize that word
• Determine if the object is on the table,Shadows that focus on lighting are required

The derivation of the one-term function through the gray function,得到极值


对x求导,就是yThe line of direction is a little more obvious

Use the modulo value of the gradient,reflect edge information

非极大值抑制

If it is not the maximum value, it will be deleted for you,Only the maximum value is kept

High threshold first,Then low threshold,
Only and high threshold adjacent low threshold （When the two are connected into a complete line）才会被保留

canny

纹理

Don't read this part（感觉用不着）

Use convolution kernel groups

I'm a little confused about this smooth feeling

反思

Too little time to study..Too much time to play（2:6）