The first debate is on "Will contrast enhancement improve supervised classification accuracy?"
Preamble:
Analysts perform contrast enhancement to better visualize their objects of interest. So would this enhancement improve the between-class separability?
Scenario 1:
The analyst contrast stretches the image and chooses the training pixels. He burns the LUT (look up table) to create a new image. He uses the same training pixels for classification in the original image and on the stretched image.
Scenario 2:
The analyst uses a different set of traning pixels on both the images.
I would like to know your opinion on the results of the two scenarios. Which one would you consider better and the reason.
3 comments:
i will probably go for scenario 2 - Contrast enhancement(acc to me) is mainly for better visualisation and i may not consider it from the accuracy point of view.Separate training pixels for each one will be a better option. And of course i am not sure how far that will hold. I would like to practically test n see which one is better. Wat acc to u, is the better option?
The second scenario certainly seems to be the best and in some ways the most logical one.
Contrast stretching is not a one-to-one mapping between the original and the stretched image. If the LUT value of a pixel becomes 163.23 (on stretching) this is rounded to 163. Theoretically speaking there is information loss. Besides if a 7-bit image is stretched to an 8-bit image, can the same training set be used for the 7-bit and 8-bit images?
Therefore I would think it only logical to use two separate training pixels for the two images.
I prefer the second scenario mostly but not for all occasions.
The result contrast enhancement varies according to the type of the function used and the results produced will vary a lot based on the function.and hence it is more apt to go for the second scenario.
but if the clasification problem in hand demands the demarcation of classes that have overlaps in the feature space and also the number of classes is small(say first order classification)i prefer the first scenario..though i have not tested it on my own i tell this based on my theoretical knowledge. if anyone have valid proof please share.
Post a Comment