Image Search

1
Image Search

• Implementation Results
• Ken Schmidt

2
Some Implementation Problems

• The server changed the path to the SQL tag
  library
• lt_at_ taglib uri"/jccrsrc/misc/jakarta-taglib/dbtag
  s.jar" prefix"sql" gt
• vs
• lt_at_ taglib uri"http//jakarta.apache.org/taglibs/
  dbtags" prefix"sql" gt

3
More Implementation Problems

• File format to upload image is MultipartRequest
  Stream, format needed for PixelGrabber is Toolkit
  Image, no conversion available
• Path problems implementing pixelGrabber
• //load the just stored file as an Image object
• Toolkit toolkit Toolkit.getDefaultToolkit()
• Image img toolkit.getImage ( new java.net.URL
• ( "http//www.mycgiserver.com/kenwschmidt/search
  /" outfile ) )
• //wait for entire image to be loaded before
  proceeding
• MediaTracker mediaTracker new MediaTracker (
  new Frame ( ) )
• mediaTracker.addImage ( img, 0 )
• mediaTracker.waitForID ( 0 )
• // out.println ("ltpgtwidth of Image
  "img.getWidth(null)
• // "lt/pgtltpgtheight of Image "img.getHeight(nu
  ll) "lt/pgt")
• //put the Image pixels in the pixels array
• PixelGrabber pg new PixelGrabber(img, 0, 0,
  width, height, pixels, 0, width)
• try
• pg.grabPixels()

4
Calculating Pixel Values for DB

• Run through pixel array twice, once to obtain the
  mean
• Second time to calculate moments
• redSum2 Math.pow ( ( red - redMean ), 2 )
• greenSum2 Math.pow ( ( green- greenMean ), 2
  )
• blueSum2 Math.pow ( ( blue-blueMean ), 2 )
• redSum3 Math.pow ( ( red - redMean ), 3 )
• greenSum3 Math.pow ( ( green- greenMean ), 3
  )
• blueSum3 Math.pow ( ( blue- blueMean ), 3 )
• redSum4 Math.pow ( ( red - redMean ), 4 )
• greenSum4 Math.pow ( ( green- greenMean ), 4
  )
• blueSum4 Math.pow ( ( blue- blueMean ), 4 )
• Then enter all values into DB

5
Now to Search the DB

• Three steps
• View all Items in the DB, Select one to match
• Enter search criteria
• Percent deviation from mean and moment values
• Which DB values to search
• Retrieve images whose mean or moment values are
  within the limits
• Calculate the upper and lower limits of each
  value (1 ? ( deviation /100 ) )
• Select images within the limits
• select from search where OVERALLMEAN BETWEEN
  ltoverallMeanDowngt AND ltoverallMeanUpgt

6
Some ResultsWhich Search Methods Worked?
CM-Color Mean GM-Grayscale Mean C2-Color Second
Moment G2-Grayscale Second Moment
7
More ResultsWhich Search Methods Worked?
8
More ResultsWhich Search Methods Worked?
9
More ResultsWhich Search Methods Worked?
10
Conclusions

• In each case, taking pixels grouped by color
  seemed to produce better results than taking all
  colors together
• The third moment did not work well, because it
  can produce both positive and negative values and
  I only allowed a deviation maximum of 100 of the
  value of the search image
• The Cambridge Lab used color and texture without
  image segmentation in combination, I used them
  separately so that I could see which method
  produced the best results.