The problem of separating structures from the rest of a USGS Topographic Quadrangle became necessary because of the need to find rural residences.  Because of USGS standards and image / palette compression, several features share the same color in the map.  This fact makes tools like Adobe Photoshop useless, because they search by color.  We therefore needed a tool to search for patterns.  FullPixelSearch is a tool (for Macintosh) that met some of our needs.

The process of extracting structures involved more applications than "should" be necessary.  I ended up using Adobe Photoshop to pre-process the Digital Quads (off of CDs from USGS), then FullPixelSearch to "find" the structures, then Debabelizer from Equilibrium to translate the output file into something Photoshop could read, and finally Photoshop to remove everything except what FullPixelSearch found and save back to a .TIF file format for opening in ARC (Info or View).
 

Pre process

Because the files from the USGS were in index color .TIF format, and FullPixelSearch (FPS) only accurately processes rgb .pict, there had to be a bit of file translation before starting.  Photoshop v.5 has scripting capabilities (referred to as actions).  I "wrote" an action that opens the image and saves as rgb pict.  This action can be a batch process, so multiple files could be processed.

FullPixelSearch

For an initial test to see how useful FPS would be, I processed 23 Quads for the Marion County area. There had been problems getting FPS to batch process images (working now), so I did them all individually.  Each image is between 8-24MB and takes about 5 minutes to open, process, and save. The resulting image is similar to the original image except that "found" areas have been highlighted with solid red (a unique color from other red in original Quad).

Post process

After FullPixel Search had done it's thing, the structures had to be separated from the rest of the image. Unfortunately FPS doesn't write out a .pict file that Photoshop can read, so I used Debabelizer to open all the images from FPS and re-save them as .pict.

Once I had images that Photoshop could read, I used the Actions in Photoshop.  The "set" of actions I "wrote" involves searching for the unique color that FPS added to the image, and selecting it.  Then Photoshop fills the selection with black (structures) and inverts the selected area and fills it with white (background).  This action can be a batch process also, so I ran this on all the output files from FPS.  And finally back to Debabelizer to re-save the images as "packed" grayscale .TIFs.

The resulting image

The results from FullPixelSearch are much better than any other tool available to us. See example1  That isn't saying that they are perfect.  Because some roads are solid areas of color, FPS can find the 'pattern" of pixels within the road.  This results in FPS highlighting large stretches of road.  Text has similar problems. See example2.  There are also some structures that are represented by a box with a black outline and a white interior.  It has not been possible to extract these structures.  There are also areas where individual structures are not mapped, so FPS can't do much with these either.

Additional Post Processing

We were hoping to use some features of ArcInfo to remove the roads and maybe large text, however it became necessary to use Photoshop to erase the rest of the text by hand, so I made 8 CDs of 18-20 quads each and passed them out. It took a few weeks for everyone to finish (working on the quads when they had time), but the results are pretty good, and seem to be filling a gap in our data for rural areas.