I have a bunch of methods for image thresholding and, I created a reference image without using any ground truth from the existing methods (Say the mean of all thresholded value of an image is the reference image). However when I try to plot the ROC curve with X- axis False positive rate (FPR) and Y-axis True positive rate (TPR), it is shown that for every method of thresholding the plot is coming on either X- axis or Y- axis. That means either my false positive(FP) or false negative (FN) value is becoming zero. Which gives the Area Under the Curve (AUC) is always 1, which is quite impossible. The reference image threshold value is within the range of existing methods. Hence my question is -
1 - Is it always true that if a point is within the cluster, then it's FP and FN value will be zero?
2- If yes, then why? if no, then why it happening in this scenario ?
3- Is there any Mathematical tool to support this statement?
Thans and Regards
Susant Kumar Panigrahi
NIT, Rourkela.
I suggest you to get a look to this paper, and use DET plots instead of ROC "A. Martin et al., "The DET Curve in Assessment of Detection Task Performance", Proc. Eurospeech '97, v. 4, pp. 1895-1898"
you can use a MATLAB toolbox available here: http://www.itl.nist.gov/iad/mig//tools/ (detware)
@Vahid Nouri, I think you can use 'cftool' (Curve fitting tool) to find out your ROC curves, by taking average at each point of the plot you can get the mean values and can plot it easily.
Regards,
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing