At present, the research about fast algorithm for multi-threshold image segmentation is mostly focused on bionic algorithm. Many of bionic algorithms, such as particle swarm optimization algorithm (PSO), artificial bee colony algorithm (ABC), bat algorithm (BA), firefly algorithm (FA) and grey wolf optimization algorithm (GWO), have achieved good results and are being further improved. In this paper, a gradient-based method different from bionic algorithm is proposed to search optimal multilevel thresholds. For most realistic images, the objective functions in mostly-used multi-threshold methods, such as Kapur’s entropy and Otsu function, have good convexity. On this basis, gradient descent method can be used to solve the optimization problem more simply and directly. Based on gradient descent method, our method is tailored for discrete objection function, and the process of searching optimal threshold is divided into two stages, which makes algorithm convergence efficient and accurate. The multi-thresholding experiments of above mentioned algorithms have been conducted, whose results show that the gradient-based method is efficient in computation and has equal or even better performance of segmentation than other algorithms.