Wednesday, March 19, 2003
( 8:31 AM ) Matt
A New Photomosaic
From a distance it looks like a regular image, but looking closer, you realize that it is made up of many smaller images. In a regular photomosaic, the smaller images are complete, rectangular photos arranged in a grid. What I propose is to do away with the grid and rectaliarity of the photos. Instead, arbitrary portions of the source images will be used, and they will be carefully chosen so as to blend into eachother.
I ran this idea by Jesse, and he pointed out that there are some neat properties of photomosaics that you don't want to lose. That reminds me of something one of my composition profesors once said. A student asked, "Why do we have to write our piece in this form? Why can't we write it however we want?" To which he replied, "The form gives a framework for you to work in. There are still pleanty of possibilities for the piece." By eliminating the restriction of the grid, I eliminate the framework which defines a classic photomosaic. This is indeed a loss. For example, there is something beautiful which comes from the ristricted form of a haiku that can't be captured in free-form poetry. Never the less, I hope that the removal of the grid restriction from my photomosaic doesn't hinder the quality of the results.
One pitfall which opens when the grid restriction is lifted is that edges desired in the target image may be created by the boundary of two reduced images. I think this is something Jesse was hinting at with his comment.
If you've read my previous entries, I am sure you can guess what my solution to this problem is -- Wavelets. Wavelets have the convinient ability to capture both spacial and frequency information. The spacial information tells us something about an image at any specific location. The frequency information tells us about the texture at that location. A wavelet moment is all the frequency information for a single location. Not only does this completly describe the color at that location, but it also describes the texture at that location and has some information about surrounding locations.
If we search our image database for a match with the wavelet moment for a location in our desired target image, we will find an image which not only matches the color of our target image, but also the texture in that area. Adjacent portions of the reduced image may be used as long as the wavelet moments are sufficiently similar to the adjacent wavelet moments of the desired target image.# -
Comments: Post a Comment