Image thresholding through nonextensive entropies and long-range correlation

In recent years, many image thresholding techniques have emerged involving entropy measures with the related long-range and short-range correlation properties. However, despite the segmentation capabilities demonstrated by those methods, we have noticed limitations in dealing with images with local long-range correlation in the foreground and background. In order to address this issue, in this paper, we propose a combination of two approaches, the first one that applies the Tsallis and Shannon entropies while the second one uses the Masi entropy as the information measure. Such a combination leads to a thresholding criterion based on Tsallis and Masi entropies, providing an improved long-range correlation image thresholding method. Besides, differently from the others, the novel technique works with two entropic parameters instead of just one, which improves the technique’s capabilities to fit the specific requirements of the applications. In the computational experiments, the quantitative evaluation of the segmentation is performed using infrared, Non-Destructive Testing images, the public Berkeley Segmentation Dataset (BSDS500), together with four error metrics computed through the ground-truth segmentation and the obtained results. The proposed method outperforms the competing approaches for infrared and non-destructive images. In the case of BSDS500, we get the second best results. For benchmark images without ground-truth segmentation, the visual analysis shows that the proposal is competitive concerning counterpart techniques.