The first image process I did is build 2D convolutions filter using finite difference operator:
1. Find the dx and dy in order to get the partial derivative of the image by
2. We compute the graient magnitude of the image by summing the squaring of the partial derivatives and then take the square root of it.
3. We take a threshold and make higher to 1 and lower to 0.
Sample Outputs:
In this part, we applied a gaussian filter on the image, then do the finite difference operator on the result blurred image.
Compare with the regular finite difference operator, the result with DoG seems to have less noise due to the blurred process, since gaussian filter is acted as a low pass filter, it will block many high frequency noises.
In addition, the edge become thick, which is also a result from blurring.
Sample Outputs:
Then, since the convolution is commutative, we can aternatively create the same result by first doing the difference operator on the gaussian filter, and then convolve the filters with the image. As we can see on the two result images create by double and single DoG respectively, the result are the same and can both help us to get rid of noises.
Sample Outputs:
When we have a blurry image, one technique that we can use to make the image clear is to utilizing the sharpening technique.
In order to do so:
1. We construct a low pass filter and perform it on the image to create a blurred image(low frequency).
2. Use the original image minus the blurred image to get the high frequency part of the image.
3. Finally, we add the orginal image with the high frequency part times alpha(parameter defined) to construct a sharpened image.
Next, I tried to create hybrid images by combining low frequency of one image with the high frequency of the other. Thus, if we look at the image close, we will se the high frequency image more. If we get far away from the image, we will see the low frequency image more.
Bells & Whistles
Here are some examples with colored:
Potential Reason of failure:
The low frequency(basketball) has bold edges which will still remain having high frequency after doing low-pass filter and thus remains visiable in the resulting image.
Using color to enhance the effect works better on high-frequency component and it becomes clearer when watching close.
In order to create a great blending between two images, we need to first implement the Gaussian and Laplacian Stacks.
We can implemenet them by extracting the low and high frequency components of an image, then do the same on the blurred version.
Aurora at Fuji
Fuji
Aurora
Mask
Aurora at Fuji
Spongebob Riding Horse
Spongebob
Horse
Mask
Spongebob Riding Horse
Reallife Snowball
Snowball
Pet Rabbit
Mask
Reallife Snowball
Spongebob with Minions' Eyes
Minions
Spongebob
Mask
Spongebob with Minions' Eyes
Minions at Berkeley
Minions
Berkeley
Mask
Minions at Berkeley
In this project, I utilized filters and frequencies and applied them on various of image processing techniques including sharpening, bluring, making hybrid images and blending two images.
I enjoyed the blending process with implementation of multiresulution pyramid.
It is my first time I learned how frequency is used to create results that are blurred or sharpened.