Self noise and contrast controlled thinning of gray images

• Review of thinning methods including gray parametric thinning.
• Ability to separate self-image noise and contrast impact on parametric thinning setting.
• Standardization of the setting through the use of statistical test framework.
• Assessment protocol of gray thinning methods based on mandatory properties.
• High performance of our method regarding homotopy and extremity preservation properties.

Homotopic grayscale thinning leads to bushy skeleton when applied on noisy images. One way to reduce this phenomenon is the use of the parametric thinning approach. It consists in relaxing the initial constraint by lowering low-contrast crests, peaks and ends, according to a manually selected parameter and under the constraint of ascendant gray level processing. In this work, we propose to control the thinning parameter by considering the lowering decision as a hypothesis testing of a statistical framework. A unitary hypothesis test based on the minimum test statistic is used for the elimination of noise-related peaks and extremities, while a fusion of multiple tests is performed for the insignificant crest lowering decision. This statistical control is first detailed under the assumption of additive Gaussian noise and then, is generalized for noise distributions with known pivotal quantity. The proposed statistical control leads to a local adjustment and a standardization of the parametric thinning process that depends on both the test significance level which is linked to image contrast and to noise standard deviation. The proposed method is tested on synthetic and real images, and compared to two skeletonization methods with proven efficiency.