Project 1 NOISE Instructions

[rzant]

Within the article by Leon Harmon, the concept that interested me most was the limit to which an image could be distorted while maintaining recognizable elements. Within their study, it turned out that the minimum resolution that allowed for identification of the image was 16 by 16 squares. What is more fascinating is the fact that distorting the image to a greater extent, by adding either blur or noise, can make the image more recognizable. This additive distortion process differs from the subtractive nature of lowering the resolution, but it still seems counterintuitive that an addition of random information may aid in image recognition. However, Harmon explains that the addition of noise or blur may act as a low pass filter, eliminating all high frequencies that are created by a degrading of image resolution. It is then easier to reconstruct missing information based on purely low frequencies and without high frequencies to distract and mislead the eye.
Manipulating images
In producing my images, I wanted to play with the threshold of minimum information that an image retains from its original copy. Much of my manipulation involved reducing the image to as low as 4x4 pixels, and then stretching that information to as large as 4000x4000 in size. I repeated this step many times on each image, also decreasing either the width or the height to a low amount of pixels while the opposing parameter retained a high amount of pixels. Needless to say, these manipulations would cause the image to change drastically. In 3 of the five images, there was a point at which the image appeared as a uniform color (either brown or grey). However, true to the Harmon article, the addition of noise and blur would often cause recognizable elements found within the original image to reappear in the distorted version. Thus, there are ways to degrade the image and rid it of almost all of its information, but bring back some of its original information with the addition of noise and blur. I thought that these processes of simultaneous destruction and recreation were fascinating. It was also interesting to see completely new forms appear that were not a part of the original image and to attempt to derive some type of structure and form from the addition of random information.

1. Squares

Original size: 500x333

Size: 500x3333 (bilinear)
Noise: Gaussian, 9.5%
Size: 100x100 (bicubic)
Blur: Gaussian, 3.8 pixel radius
Noise: Gaussian, 9.5%
Size: 100x1 (bicubic)
Size: 100x1000 (bicubic)
Size: 5x5 (bicubic sharper)
Size: 1000x1000 (nearest neighbor)
Noise: Gaussian, 9.5%
Noise: Gaussian, 9.5%
Noise: Gaussian, 9.5%
Noise: Gaussian, 9.5%
Size: 333x555 (bicubic)
Size: 2000x3942 (bicubic)
Blur: Gaussian, 32.8 pixel radius
Noise: Gaussian, 345.9%
Size: 4x7 (bicubic)
Size: 3000x4218 (bicubic)
Levels: Composite channel, 109,247
Levels: Gamma, 2.33
Exposure: 1.22
Offset: -0.182
Gamma Correction: 0.9
Size: 1000x1000

2. Joaquin

Original size: 445x424

Lightness: 39
Noise: Gaussian, 400%
Size: 1x424 (bicubic)
Size: 1080x424 (bicubic)
Rotate Canvas: -90°
Add Noise: Gaussian, 51.9%
Image Size: 1x424 (bicubic)
Image Size: 1080x424 (bicubic)
Image Size: 1080x1 (bicubic)
Add Noise: Gaussian, 51.9%
Image Size: 1080x1080 (bicubic)
Image Size: 10x1080 (bicubic)
Image Size: 10x10 (bicubic)
Image Size: 100x100(bicubic)
Blur: Gaussian, 3.2 pixel radius
Add Noise: Gaussian, 51.9%
Size: 20x100 (bicubic)
Size: 4x4 (bicubic)
Add Noise: Gaussian, 51.9%
Size: 4000x4000 (bicubic)
Rotate Canvas: 90°
Size: 40x400 (bicubic)
Add Noise: Gaussian, 51.9%
Size: 4000x400 (bicubic)
Size: 4000x4000 (bicubic)
Blur: Gaussian, 1.6 pixel radius
Blur: Gaussian, 1.6 pixel radius
Add Noise: Gaussian, 51.9%
Size: 5.5x4000 (bicubic)
Size: 55x4000 (bicubic)
Size: 5x5 (bicubic)
Size: 1000x1000 (bicubic)

3. Structure

Original size: 3648x2376

Project 1: Noise and Form
Add Noise: Gaussian, 51.9%
Size: 500x500 (nearest neighbor)
Size: 3000x3000 (bicubic)
Size: 30x50 (bicubic)
Blur: Gaussian, 0.9 pixel radius
Levels: composit channel, 55,188
Size: 4000x4000 (bicubic)
Flip Canvas: Vertical
Add Noise: Gaussian, 70.7%
Size: 4000x3000 (bicubic sharper)
Size: 40x30 (bicubic)
Add Noise: Gaussian, 70.7%
Size: 20x30 (bicubic)
Levels: composit channel, 56,199
Size: 5x30 (bicubic)
Size: 50x30 (bicubic)
Size: 50x3 (bicubic)
Size: 50x39 (bicubic)
Size: 3000x3000 (nearest neighbor)
Levels: composit channel, 29,184
Gamma: 0.9


4. Futbol

Original size: 416x594

Size: 1000x1900 (nearest neighbor)
Levels: composite channel, 61,239
Gamma: 0.73
Size: 100x1900 (bilinear)
Noise: Gaussian, 2.5%
Noise: Gaussian, 2.5%
Size: 100x190 (nearest neighbor)
Size: 90x90 (nearest neighbor)
Levels: composite channel, 43,259
Gamma:1.33
Size: 30x30 (nearest neighbor)
Size: 3000x3000 (nearest neighbor)
Size: 3000x100 (nearest neighbor)
Size: 4000x1900 (bilinear)
Size: 30x1900 (bicubic)
Size: 3000x1900 (bicubic)
Size: 25x3000 (bicubic)
Size: 2500x3000 (nearest neighbor)
Noise: Gaussian, 55.6%
Size: 25x30 (bilinear)
Levels: composite channel, 54,203
Gamma: 1.44
Size: 4000x3000 (nearest neighbor)
Rotate: 90°
Size: 30x3000 (nearest neighbor)
Size: 2000x3000 (bicubic)
Levels: composite channel, 44,243
Gamma: 1.3
Size: 3000x5000 (bicubic smoother)
Size: 3000x50 (bicubic)
Size: 3000x5000 (nearest neighbor)
Noise: Gaussian, 2.5%


5. Riot

Original size: 450x298

Harmon article
Add Noise: Gaussian, 28.3%
Size: 4000x20 (nearest neighbor)
Size: 500x20 (bicubic)
Size: 4000x20 (bicubic)
Size: 4000x4000 (nearest neighbor)
Levels: composite channel, 0,158
Gamma: 0.67
Size: 40x20 (bicubic smoother)
Size: 100x100 (nearest neighbor)
Size: 4x20 (bicubic)
Blur: Gaussian, 9.5 pixel radius
Add Noise: Gaussian, 28.3%
Size: 1000x100 (bilinear)
Size: 999x6 (bilinear)
Size: 1000x1000 (nearest neighbor)
Add Noise: Gaussian, 28.3%
Size: 50x40 (bicubic)
Add Noise: Gaussian, 28.3%
Size: 3000x3000 (nearest neighbor)
Size: 3x3000 (nearest neighbor)
Size: 3000x3000 (bilinear)
Size: 30x30 (bicubic)
Add Noise: Gaussian, 28.3%
Size: 5x30 (bicubic)
Add Noise: Gaussian, 28.3%

[leighdodson]

My image manipulation was inspired by the Harmon article, particularly his experiments with recognition and reduction. When researching into the drawing of faces used by police to illustrate criminals from witness reports, he states that the first initial drawings were strikingly similar to the actual appearance, except for the addition of extraneous details not specified by the witness. I linked these details to the addition of noise in Photoshop- it is the random addition of information that is not there. I also felt that when reducing and expanding that photo, Photoshop has to stretch few pixels into many, adding in information where there is not enough for the new, larger size. The sharpen tool, as well, forces the computer, like the artists of criminal investigators, to make and assume information in order to become more accurate.

[leighdodson]

Images continued

[leighdodson]

Image 1.
Waste

minimize to 1% vertically; 10% horizontally
return to original size
add noise (25%)
invert
pixelate-fragment
blur
smart sharpen

Image 2.
Mess

Add noise (15%)
blur
compress to 1% vertical and 1% horizontal
return to original size
blur
smart sharpen
increase brightness and contrast to 90%


Image 3.
Time Machine

Compress to 5% vertically and horizontally
Return to original size
Pixelate-Fragment
Smart Sharpen
Fragment
Fragment
Fragment
Smart Sharpen
Invert
Increase Contrast to 80%


Image 4.
New Orleans
Compress to 1% vertically and 1% horizontally
Return to original size
Blur
Blur
Invert
Increase contrast to 50%
Pixelate-fragment
Smart sharpen
Blur