Maximum similarity thresholding

• This research provides a maximum-image-similarity view of Otsu method.
• It constructs a flexible maximum similarity thresholding (MST) framework.
• It proposes a new thresholding method based on the MST framework.
• New method can deal with images with strongly unbalanced class sizes.
• It is robust to thresholding images with Gaussian or non-Gaussian mixture distributions.

Otsu method is one of the most popular image thresholding methods. The segmentation results of Otsu method are in general acceptable for the gray level images with bimodal histogram patterns that can be approximated with mixture Gaussian modal. However, it is difficult for Otsu method to determine the reliable thresholds for the images with mixture non-Gaussian modal, such as mixture Rayleigh modal, mixture extreme value modal, mixture Beta modal, mixture uniform modal, comb-like modal. In order to determine automatically the robust and optimum thresholds for the images with various histogram patterns, this paper proposes a new global thresholding method based on a maximum-image-similarity idea. The idea is inspired by analyzing the relationship between Otsu method and Pearson correlation coefficient (PCC), which provides a novel interpretation of Otsu method from the perspective of maximizing image similarity. It is then natural to construct a maximum similarity thresholding (MST) framework by generalizing Otsu method with the maximum-image-similarity concept. As an example, a novel MST method is directly designed according to this framework, and its robustness and effectiveness are confirmed by the experimental results on 41 synthetic images and 86 real world images with various histogram shapes. Its extension to multilevel thresholding case is also discussed briefly.

Figure optionsDownload as PowerPoint slide