-> User's Manual -> Color Segmenter

Related Pages
Document Handling
White paper on document handling

Segmenter Settings
How to configure the segmenter

Contents of This Page

Noise Sources - Where noise comes from

Segmentation and Restoration - On segmentation and image restoration

Noise Sources

Transition Distortion : When an image is electronically captured, it is usually significantly distorted. Different pictures have different types of distortion ("noise"). For example, a picture that was encoded into a television signal such as NTSC (National Television Standards Committee) is going to have a different kind of noise than a picture that comes from a scanner.

Pac-n-Zoom® is like a set of glasses for your computer. It can correct the digital picture by removing much of the distortion, but correct settings (like lens prescriptions) will be needed to get the best picture. In this paper, we will look at how noise gets into a picture and how Pac-n-Zoom can restore some of the original picture.

Blurring is the worst noise in almost all digital images. It has two primary sources.

1. Transition Pixels: Let's suppose a round dot is printed on a piece of paper as shown in the "Transistion Distortion" image. If we were to scan the dot into the computer, we might sample the dot as the sample grid row shows.

Transition Distortion

Transition Distortion The outside and middle pixels are all white or black respectively, but there are two edge pixels that are part white and part black as shown in the ratio grid row.

When converted to gray scale in an array, we have several different ratios; so two colors (black and white) have been changed to black, white, and 3 shades of gray. Without perfect symetry, we could get 12 or more shades of gray.

Pac-n-Zoom can "sweep" some of the gray out of the white and over to the black without changing the average color of the picture. Pac-n-Zoom can be set to restore the picture to the two original colors, but the average color will be changed slightly.

2. Optical Aberrations: A very small (about 1/4th of a millionth of a square inch per sensing cell or pixel) optical array is used to sample the dot. The optical image cast upon the array will add to the transition pixels for several reasons. Lets look at two of the more common problems.

a. Color: Lens are ground to a specific wavelength of light. White is not a primary color (it has a mix of other wavelengths). Therefore chromatic abberrations (distortions due to different wavelengths of light) will occur.

b. Focus: A lens is only focused perfectly at a specific depth of field. In a scanner, the optics can be set to minimized this effect, but photographs often have very big problems in this area.

Since optical abberations blur the image cast onto the sensing grid, they cause the number of transition pixels to magnified. Pac-n-Zoom can be set to sweep different amounts of transition pixels.

Color Flutter:

Another type of noise (we will call flutter) multiplies the number of artifacts in a picture by creating slight changes in a specific color. Flutter can occur from electronic sources such as clock coupling, Johnson noise, dark current, beta mismatches, and a number of other ways. For example, flutter also occurs from a lack of homogeneity in the image light source.

In another application, a digital speedometer has flutter when the car is going 54.5 miles per hour, and the speedometer vacillates between 54 and 55.

To eliminate flutter, we could simply require a 2 mile per hour change before a new speed was displayed. This would be noticed on a speedometer, but it would not be noticed in a 24 bit continuous tone color picture. More than 90% of the people we tested could see the difference between 4,096 (4 bits per primary color with 3 primary colors) and 32,768 (5 bits) colors, but only about 18% of the people could see the difference between 32,768 and 262,144 (6 bits) colors. From these numbers, we can conclude 6 bits of color on 3 primary colors is probably enough.

Pac-n-Zoom can be set to eliminate most flutter by requiring the color jump more than one level (7 bits of color are left). To prevent using resources on invisible color variations, the flutter setting is adjustable.

Orphan Pixels:

Another type of noise (we A picture's value is derived from the information in the picture. In other words, if no information can be extracted from the picture, it has very little value.

To extract information from a picture, a group of pixels must be combined together into a constellation. Therefore a single pixel that can not be combined with any other pixel (an orphaned pixel) does not add value to a picture, and it should be eliminated.

Orphan pixel can be caused by dust specs on the optics, dead pixels in the camera element, or by large electronic noise spikes. Small artifacts with high contrast are often a source of orphan pixels, but even in these cases, the orphan pixels usually detract from the picture.

If the picture was painted on the computer by an artist, the artist might use the spray tool to create a shading and therefore orphans. The artist knows that your eye will average the orphan into the adjacent pixels. When Pac-n-Zoom eliminates the orphan, it averages it into the blob of an adjacent pixel. Therefore, the same artist created shading will still exist after the orphan elimination.

All of Pac-n-Zoom's minimum artifact settings eliminate orphan pixels.

Segmentation and Restoration

Restoration : Pac-n-Zoom® can be set to eliminate transition and flutter noise, and Pac-n-Zoom always eliminates orphans by averaging them into a blob of an adjacent pixel.

These are three common types of noises in digital images. By removing these three types of noise, the clarity, the contrast and the compressibility of the picture are all enhanced.

Segmentation : Before intelligence can be extracted from a picture, a group of pixels have to be aggregated together and recognized by an intelligent extractor (usually a human being). When the picture is made more simple, it compresses better. Furthermore, it is usually more pleasing to the human being because the human being does not have to subconsciously aggregate artifacts that are inhibiting intelligent extraction. Pac-n-Zoom improves the picture appeal and compressibility by aggregrating the blobs; so the human does not have to work as hard.

Back - Back to Pac-n-Zoom diagram

© 2004 - 2011 Accelerated I/O, Inc, All rights reserved.