Study On Methods Of Thresholding Image Segmentation Based On Tsallis Entropy

Image segmentation, which extracts interested objects from image, is an important research topic in low-level computer vision. Image segmentation is a crucial pre-processing step for image analysis and understanding. The quality of image segmentation has great effect on image analysis and understanding, image recognition. In the past twenty years, researchers have extensively studied thresholding method for image segmentation based on Shannon entropy and proposed many methods. However, thresholding methods based on Shannon entropy ignore the relationship between the object and background due to the extensiveness of Shannon entropy, which may cause inaccurate segmentation result in some situations.Tsallis entropy is a possible extension of Shannon entropy, in Information Theory, which can describle more complex statistical properties (long-range interactions, long time memory and fractal-type structures) than the Shannon entropy. For the non-extensive(pseudo-additive) of Tsallis entropy, we apply it in the image segmentation area, and consider the relationship between the object and background probability which improves the accuracy of segmentation result. The paper fulfills the following work:First, the minimum Tsallis-cross entropy thresholding method is proposed. The proposed method measures the information difference between object and background using Tsallis relative entropy. Since the information difference measured by Tsallis relative entropy considers the relationship between object and background furthermore, the segmentation results obtained by the proposed method improves distinctly.Second, the minimum Tsallis relative entropy thesholding method was extended to two-dimensional case. Two-dimensional minimum Tsallis relative entropy thresholding method is robust to noise because it considers not only the gray level information but also the spatial relationship of pixels. Particle swarm optimization is used to search the optimal threshold vector.Finally, fuzzy Tsallis entropy thresholding method is proposed by considering the fact that many fuzziness exist in image processing. The proposed method transforms the image into fuzzy domain to obtain a fuzzy 2-partition of the image using parameterized membership function. Then, Tsallis entropy of fuzzy 2-partition is defined. The ideal threshold are obtained by searching an optimal combination of parameter such that the Tsallis entropy of the fuzzy 2-partition is maximized. Experimental results show that the proposed method can get satisfactory result.