An Image Segmentation Method by Multi-scale Local Thresholding Based on Class Uncertainty Theory
Likai Zhou, Yuyang Jiang, Guoyuan Liang, Xinyu Wu, Jiafeng Zhu, Huaikun Xiang · 2019
Image segmentation is one of the most important core technologies in the field of image processing and computer vision, which has lots of applications specifically in medical image analysis. Due to medical imaging mechanism, medical images usually suffer from heavy noises, uneven intensity distribution and fuzzy boundaries between biological tissues. That makes classical segmentation methods based on global or local thresholding difficult to get accurate results. This paper proposes an image segmentation method by multi-scale local thresholding based on class uncertainty theory. Firstly, the original image is divided into a set of sub-regions with different scales by using a multi-layer pyramid structure. Secondly, the energy function with inequality constraints is constructed based on class uncertainty and region uniformity. Then by an iterative process, the optimal local threshold for each sub-region are calculated by an optimization algorithm for each layer until the final optimal threshold mask is determined for image segmentation. Experimental results verify that the proposed method can effectively eliminate the interferences caused by image noises, intensity unevenness and fuzzy boundaries, and preserve details of object structure while segmenting the object from the background. Segmentation results demonstrate the superior performance of the proposed method by comparison with supervised range-constrained thresholding methods (RCOtsu), classical global/local Otsu method, and the minimization of homogeneity- and uncertainty-based energy method (MHUE).