Ideas to improve per-pixel classification

Hello,

I'm working on a pixel classification of an image. I have 4 classes and they are using RGB coefficients. I have one problem using SVM classification, there is 2 of the classes wich are similar. In the confusion matrix, I have an error of 40% between that 2 classes. They have pretty high similar colors. Does anyone have an idea or advice for improving the accuracy of the classification? For example adding more predictive or adding more contrast between these 2 problematic classes

Thank you

Answer 1

[Image Analyst]

What is the basis for your classification? Color? Are you comparing your image to 4 known reference colors? Then try using Delta E or LDA.

[ben dp]

well yes I have RGB image and so I'm using the RGB coefficients for SVM classifier. And I have to use SVM so I cna't use Delta E or LDA> Do you have any other suggestion? Thank you!

[Image Analyst]

Remember you've chosen not to share your image, so I'm just guessing here. It could be that the colors are actually pretty close together and be difficult to separate. You can use the function colorcloud() to see the 3-D color gamut and see your clusters, if indeed you even have any. Many natural scene images don't have clearly defined clusters and the "clusters" blend from one to the other with no clear cut dividing line. Then you might have to try other things in addition to color classification to segment out your objects, like perhaps based on size or shape or texture of the regions/objects. Perhaps PCA could help get you a better coordinate system for extending the length of the "dual cluster" region and allow you to split it better. I attach a demo of PCA (requires stats toolbox).