Abstract: The segmentation of metallographic images plays a key role in grain grading, but it is difficult to extract grains accurately using the traditional Chan-Vese (CV) model. To segment metallographic images more accurately, a metallographic image segmentation method based on an improved CV model was proposed. First, the level set function was initialized, and its reciprocal Canberra distance from inside and outside the curve was calculated. Then, these distances were used as weight coefficients of the fitting centers to restrain the influence of noise points on their accuracy. In addition, adding exponential entropy to adjust the energy inside and outside the curve reduces the influence of the fixed energy weight on the evolution of the curve. Lastly, to accelerate the convergence of the model, a distance-regularized term was introduced to avoid re-initialization of the level set function. The experimental results show that, compared with the traditional CV model, the geodesic active contour model, the distance-regularized level set evolution model, and the bias level correction level set model, the segmentation of the metallographic images based on the proposed model is more accurate and efficient, and the proposed model has better convergence.