Seeking multi-thresholds directly from support vectors for image segmentation

Threshold selection is an important topic and also a critical preprocessing step for image analysis, pattern recognition and computer vision. In this letter, a novel automatic image thresholding approach only from the support vectors is proposed. It first fits the 1D histogram of a given image by support vector regression (SVR) to obtain all boundary support vectors and then sifts automatically so-needed (multi-) threshold values directly from the support vectors rather than the optimized extrema of the fitted histogram in which finding the extrema is, in general, difficult. The proposed approach is not only computationally efficient but also does not require prior assumptions whatsoever to be made about the image (type, features, contents, stochastic model, etc.). Such an algorithm is most useful for applications that are supposed to work with different (and possibly initially unknown) types of images. The experimental results demonstrate that the proposed approach can select the thresholds automatically and effectively, and the resulting images can preserve the main features of the components of the original images very well.