Level set framework based on local scalable Gaussian distribution and adaptive-scale operator for accurate image segmentation and correction

Abstract

Active contour model is a popular image segmentation method, but its current development is still limited by the problem of uneven image intensity. This paper proposes a new active contour model based on level set framework using local scalable Gaussian distribution (LSGD) and adaptive-scale operator for accurate image segmentation and correction, and uses the split Bregman method to minimize the energy functional. We construct a fitting energy with local Gaussian distribution characteristics, and segment the image with the maximum posterior probability. Then we use an adaptive-scale operator to adjust the scale of the LSGD according to the degree of the intensity inhomogeneity. In addition, the bias field information is added to the model to achieve better robustness to unevenness and correction effects. Moreover, the energy functional is minimized by the split Bregman method to improve the efficiency of image segmentation. In order to meet the needs of more diverse image segmentation, we extend the two-phase energy functional to the multi-scale and the vector-valued energy functionals. To evaluate the effectiveness of the proposed methods, we select the composite images as well as brain MRI, melanoma and Berkeley nature images. We select the quantitative indexes to evaluate the image segmentation effect, including JS, Dice, CV, the Precision and Recall, correlation index, SSIM and MSE index. We compare our model with the classic traditional image segmentation models and U-net machine learning model. Experimental results show that the models can perform appropriate segmentation and bias field estimation, and the quantitative index values are better than other segmentation methods.