CS 194-26: Image Manipulation and Computational Photography

Overview

Play with gaussian low-pass filter and laplacian

Part 1.1: Finite Difference Operator

partical derivative of x (dx)	partical derivative of y (dy)
gradient magnitude (sqrt( dx2 + dy2 ))	edge image: binarize gradient magnitude at threshold 0.3

Part 1.2: Derivative of Gaussian (DoG) Filter

result from blurred picture(apply gaussian to picture first, then dx dy to blurred picture), have less noise and clear edges:

gradient magnitude for blurred picture;

edge image for blurred picture: binarize gradient magnitude at threshold 0.11;

gaussian filter on dx dy filter(first convolv then blur), then apply to original picture:

partical derivative of x;	partical derivative of y ;
gradient magnitude;	edge image;

The result of find gradient magnitude on gaussian filtered picture is same as first convolv dx dy with gaussian filter then apply to original picture. This corresponds with the fact that convolution is associative and commutative

Part 2.1: Image "Sharpening"

sharpen for taj

original taj	gaussian blurred taj, ksize=6, sd=1
high frequency	sharpened taj = original taj + 1*high frequency

sharpen for taj

original Angkor Wat

gaussian blurred Angkor Wat;

high frequency

sharpened Angkor Wat = original Angkor Wat + 2*high frequency

blur the sharp image and try to resharp the blur image:

original gather logo

gaussian blurred gather;

resharpened gather logo from the blurred gather

We can see from the original picture that the Gather logo is composed of many green dots. Comparing the original Gather logo with the resharpend one, we notice that the green dots is invisible from the resharpened one. Because the blurring get rid of high frequencies and details, it is impossible to get those high frequencies and details back from soly the blurred pictures.

Part 2.2: Hybrid Images

grayscale

colored

colored cat on gray face

Peacock	Folding fan
color	color peacock on gray fan

Try to hybrid zebra and leopard

Overall performance not so good to see two animals with both strong patterns to be hybrid together.

zebra	leopard	leopard	leopard
colored high frequency zebra on top of blurred leopard, cannot see the characteristic rosettes on leopard fur.	colorerd high freqeuncy leopard on blurred zebra. Gaussian ksize for zebra set to 12, still too clear stripes both from apart and from near close.	colorerd high freqeuncy leopard on blurred zebra. Gaussian ksize for zebra set to 24, better performance.