The traditional active contour model is sensitive to initial contour and noise, and is unable to extract multiple targets.To solve these problems, a novel minimum variance active contour model is proposed based on generalized likelihood ratio.The generalized likelihood ratio information is introduced to a region-based active contour model.A novel energy function is designed under the criteria of minimum variance between the target area and background area, which is then minimized by using a gradient descent method to drive the contour shrinking to object borders.Experimental results on synthetic and real images prove that the proposed model is not sensitive to the initial contour position or noise, and is adaptable to multi-target scenario.