Level Set Evolution with SOM-Based Least Squares Twin SVM for Noisy Image Segmentation

This paper presents a novel hybrid framework derived from the least squares twin support vector machine (LSTSVM) and active contour model (ACM) for noisy image segmentation. It contains a two-stage process, where the concurrent self organizing maps (SOM) are firstly employed to approximate the image intensity distributions to establish the original training sets for the LSTSVM. The training sets are then updated by adding the global region intensity means during the curve evolution. Further, the discrimination functions of the LSTSVM are embedded into the energy function of the ACM to guide the curve movement. Besides, a variable regional coefficient is designed in the energy function to enhance the noise robustness. Experiment results demonstrate that our model holds higher segmentation accuracy and more noise robustness.