| In optical science, a prism is used to break up white light into its componenty colors. The Reindeer Graphics PCA plugins do something similar for color and multichannel images. It find the principal colors that make up the image, and in so doing isolates the information in new and very useful ways. This is not just a matter of converting from RGB to Lab, CMYK, HSL, or some other standard color space. Rather, the method defines a fundamental set of basic color coordinates unique to each image. The technique used is a standard statistical method, rarely applied to image data, called Principal Component Analysis. Multichannel images 
In the process of computing the Principal Components of this set of images, a rotation in 5-space (because there are five images) with a set of new images created. Below are the most significant (73.69% - left) and the least significant (0.27% - right) channels after the transform. 
Notice that the JPEG artifacts from the set of images have all appeared in the bottom channel (right). This characteristic of PCA will prove useful for removing both pattern noise (such as JPEG and DV encoding) and random noise within images, below. Another useful trick with PCA is to place the most significant three channels into the Red, Green, and Blue channels of a standard RGB image. In this next example, Red contains the primary component (73.69%), Green holds the second component (21.67%), and Blue holds the third component (3.31%). Only a tiny fraction remains unaccounted for (1.34%) and most of that is the JPEG signature. 
This method of producing a meaningful false color image, especially when there are more than three channels is extremely useful. Stained tissue This same trick works for improving contrast on RGB images as well. Here, we take a stained tissue sample and apply a PCA transform to it to better separate the different elements of the image. 
The dark pink and light pink elements of the stained tissue have completely different statistics and therefore map into dark blue and green respectively, making it much easier to segment the image and measure the features. Noise removal Since PCA will move pattern noise to the lowest significant channel (in the RGB case this would be channel 3, in the New York City case, above, this would be channel 5), we can perform noise reduction upon that channel and then perform the inverse transform. 
By performing a simple Gaussian blur on the bottom channel, we dramatically reduce the visibility of the JPEG artefacts in this image. Other functions such as a Median filter, a Hybrid Median, Adobe's Surface Blur, and Optipix's JPEG Deblocker are all good choices for removing noise from images. Next: Image Morphology
Prev: Image Processing
Up: Fovea Pro | 