[lecture][first day] image processing

1/19(Monday)

09:30 ~ 10:30 

infrared camera has long frequency which means that the resolution is weak than visible ray.

green and gray have strong correlation. 

( blue, red have little correlation with black. )

for color image, the filter of rgb generates many sampling images, which then is merged to color image. 

high frequency in case of digital image means sudden color change. in this case, wrong sampling makes wrong color appearance.

r, g, b doesn't have totally different independent relationship.

so, you can draw rgb as three dimensional vector. but eacy value is not independent. 

RGB -> YUB conversion. 

( Y: Luminance 휘도, U, V: Chrominance, 색차. )

It intends to hide characters which human can't recognize. 

It means converting RGB to independent new values. 

HSV Color space

(H: Hue 색상, Saturation: 채도. Value: 명도)

10:30 ~ 

PSNR: Difference between original and converted one by algorithm. 

some error could be ok for human recognition but it could show big difference. 

(MSE : Mean squared error)

->

SSIM: new measure for image similarity. 

it calculates per each area. 

So, you can say if you get the good result with SSIM, your algorithm get big trust from public. 

<Image Transform>

Fourier transform ( discrete time signals )

-> DTFS ( application to DTFS to 2D image signal)

-> DCT

(extract frequency moving to horizontal X and vertical Y)

RGB->YUV(reduced to half data)->apply DCT to each 8x8 block(may select one among 16 waveforms. ) -> quantization ( only low frequency area will be remained. high frequency area will be ignored. )

(when sending DCT data, use Zigzag scan. )

when compressing, frequently used code is written with short size of bits. 

11:30~ 

그림으로 보는 최신 MPEG: 그림이 많아서 이해하기 쉬움

http://www.kyobobook.co.kr/product/detailViewKor.laf?barcode=9788970851242

H.264 AVC 비디오 압축 표준

http://www.kyobobook.co.kr/event/eventViewByMenuId.laf?dispGb=3&eventPid=1623&menuId=08

video codec is based on motion vector. 

( motion vector indicate where a previous block moves to. )

< Haar Transform (filter) >

< ART (Angular Radial Transform ) >

< Image Enhancement >

Contrast stretching: shorten some gray color section and extends some gray color section  

Histogram Equalization

You should consider whether applying equalization to overall image or partial image area. 

RGB Controlling may be dangerous so, professor recommends controlling HSV.

( RGB's 신호처리 may change color. But HSV, YUV has less risk.)

You may have to apply 잡음제거 for equalization.

( do 잡음제거 first and apply equalization because you can be sure after equalization. )

교수님 강의자료 많이 참조하였음

Digital Image Processing

Fafael C. Gonzalez

http://www.pearsonhighered.com/educator/product/Digital-Image-Processing/9780131687288.page

14:20~

<High Pass Filter >

enhance only edge area. 

<Unsharp Mask>

For removing noise, you should gathered many same images and make mean value where there's little noise. 

You can use only calculation as Laplacian operator. 

(low pass filter is applied)

then, you can enhance with high pass filter.( here you can get edge information.???) finally you add this image with the original image which is the final image. 

* you can get high pass filter as minus low pass filter image from the original image. 

교수님 email로 강의자료(pdf) 신청가능

lsh529@snu.ac.kr

<Morphology>

Erosion operator (침식연산)

dilation operator(팽창연산)

opening operator

closing operator

<Fog(Haze) Removal>

when there's same rgb ratio, it is close to gray. 

안개 없다면 dark channel은 0에 가까울 것임

안개 있으면???? -> dark channel은 

-> t(x)를 구한다.

어두울 수로 가깝다. (transmission map 거리정보를 가지고 있음. <- dark channel에 의해서 뽑는다. 안개가 많이 껴 있으면.. 즉 gray에 가까우면 원거리.)

밝은 회색을 안개가 많이 낀 것으로 간주하기 때문에 안개가 끼지 않은 회색건물이 있을 경우 error발생할 수도 있음

Nowadays, CE(Contrast enhancement) and DCP(Dark channel prior) are merged for one refined output image. 

<Edge detection>

A discontinuity/abrupt change in the intensity or color.

Get gradient magnitude of x, y.

Sobel operator is used most often. 

( more than Prewitt and Chen-Frei. )