[lecture][third day] Image processing

<Super resolution>

1. first way

Reconstruction based SR:

bacically, you should have many low resoution images. 

ex) 

POCS approach.

y = Hz + n

here, you should find y. 

first, make HR. and compare it with LR and make the HR again based on the difference. And repeat the 

previous steps. 

2. second way

Learning-based SR: learn with low and high resolution images. 

with one image, you can enhance low resolution object where high resolution same object exists. 

when you can't enhance images with just interpolation and deblurring, you can try super resolution way. 

you should consider resolution level and real time operation. 

<SIFT>

CV: so, whay feature will you search for?

with good features, you can find wanted object easily without learning scheme. 

AI, Learning: learning structure design is important. 

they study how to manage so many features from cv. 

image processing engineer's mind:

it's same as your size is as small as ant and you should recognize object around you. 

Gaussian filtering: remove noise and not change 변곡점( this is very important because 변곡점 must not be 

changed. )

<SURF>

Computer Vision A modern Approach

http://luthuli.cs.uiuc.edu/~daf/CV2E-site/cv2eindex.html

Computer Vision: Algorithms and Applications

http://lorraine.gatech.edu/computer-vision-algorithms-and-applications

<Geometric Camera Models>

Homogeneous coordinator: 카메라 원근설명시 기존 유클리드 방법에서의 변환 필요. 벡터의 스칼라값 상관 없이 같

은 것으로 인식????

goal: find how two dimension picture is projected from real three dimension object. 

( with pin hole camera or mirror)

<Three camera projections>

perspective: 

weak perspective: z가 멀어서 z고려하지 않음. 

orthographic: assume using mirror.

<2D Projective transformation>

transform 2d dimension image. 

http://darkpgmr.tistory.com/79

translation, eudlidean, similarity, affine, 3d projective. 

<Radial distortion>

you can recover the normal image from distorted image by lens

<Chromatic distortion>

<Camera calibration>

http://darkpgmr.tistory.com/32

3차원 점들이 영상에 투영된 위치를 구하거나 역으로 영상좌표로부터 3차원 공간좌표를 복원할 때에는 이러한 내부 

요인을 제거해야만 정확한 계산이 가능해집니다. 그리고 이러한 내부 요인의 파라미터 값을 구하는 과정을 카메라 

캘리브레이션이라 부릅니다.

( 아마 외부파라미터는 extrinsic camera calibration으로 따로 명칭되는 것 같기도..????)

카메라 내부 요인의 파라미터 값을 구하는 과정

(초점거리: 렌즈-이미지 센서, 주점, 비대칭계수)

<Image warping>

forward warping

inverse warping

요지는 source img에서 pixel을 회전시키고 그런데 pixel좌표 위치에 정확하게 들어가지 않을 때,

주변 pixel위치와의 거리를 계산해서 가중치를 넣어주고 주번 pixel값들에 가중치를 넣어서 평균내어 적용

2차원 영상 warping에서는 inverse warping 많이 사용.

<Panorama>

stitching(경계구하고<-알고리즘)하고 blending(blurring)한다.