Image multi-thresholding by combining the lattice Boltzmann model and a localized level set algorithm

• Authors: Souleymane Balla-Arabé;Xinbo Gao
• Venue: Neurocomputing
• Year: 2012

During the last decades, the development of high dimensional large-scale imaging devices increases the need of fast, accurate and parallelizable segmentation methods. Due to its intrinsic advantages such as its ability to handle complex shapes, the level set method (LSM) has been widely used. Nevertheless, the method is computational expensive in image segmentation, which limits its use in real-time systems and volume images segmentation. In this paper we propose an adaptive image multi-thresholding method which uses a localized level set method to detect automatically the best thresholds values from some initial given values. Instead to solve the level set equation (LSE) by using the traditional methods based on some finite difference or finite volume, we use the highly parallelizable lattice Boltzmann method (LBM). All the more, the method is faster since it is solved in histogram domain rather than the pixel domain. The time complexity is therefore considerably reduced since the number of gray levels is generally much smaller than the size of the image. The method is efficient, highly parallelizable and faster than those based on the LSM. Experiments on synthetic, real-world, medical and man-made object images demonstrate the performance of the proposed method.