Hi Ricardo. There's a couple facts missing here.

- First is, which is the objective of the thresholding procedures? Bimodal segmentation? n-modal segmentation? relevant feature discrimination?

- Second is, is it singular or multilevel thresholding? That is, do your thresholds apply on a scalar basis, or do you stablish n+1 zones on the thresholded universe?

Hi Lopez-Molina

It is bimodal segmentation. I'm searching for a single threshold.

Mr. Andre

refer this link to acquire ground truth data set of  segmentation techniques

http://www.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/

for manual segmented data set refer

http://www.wisdom.weizmann.ac.il/~vision/Seg_Evaluation_DB/index.html

for medical applications with various algorithms comparison 

http://www.bioimage.ucsb.edu/biosegmentation

dear ricardo,

well, first I do partially agree on what Hima says. Still, may I express some considerations.

a) On the BSD. The images in that dataset are highly complex. In fact, it is usually acknowledge that the (many) objects in each one cannot be discriminated by their tone, usually needing texture indicatore and other contextual features. A range of papers elaborate on this, from the PhD dissertation by Martin, 2004, to, e.g., very recent works by Mairie, Yu and Perona.

So, because the images contain several objects, and because of their complexity, I don't thing thresholding can produce meaningful results.

b) Bimodal thresholding will better be applied to super-specific environments. Medical image processing is one of those. However, depending on the characteristics of your method, I would consider one of the following three options.

b.1) If your thresholding method expects the histogram to have to heavy popuations (i.e. balanced, bimodal histogram), go for region segmentation. There's many datasets on this, specially when it comes to white-grey matter discrimination in MRI's and so.

b.2) If your thresholding methods expects the histogram to have two main populations, one being (potentially) much heavier than the other, go for linear structure discrimination. My recommendation would then be vessel segmentation from eye fundus images. There's plenty of literature and images, check out Chaudhuri (as a pioneer) and the works by Staal and co-workers, around 2002-2006.

b.3) If your thresholding method expects the histogram to have a main population, and then a non-grouped secondary, outlier population, go for feature binarization. Edges, ridges, lines... most of the processes for low-level feature extraction demand a last stage of feature binarization. Hence, any dataset for edge detection, textured region discrimination, ... any of those could be used by placing your method at the end of the image processing pipeline.

I hope that was helpful!