To solve the problems existing in the random search process of the traditional crow algorithm, such as the blindness and tendency to fall into local optimum, an improved crow search algorithm was proposed for multi-threshold image segmentation. The elite sharing strategy was adopted to make up for the blindness of the crow's position when updated. To avoid falling into local optimum, the Levy flight mechanism was introduced. Then the scale coefficients were adaptively adjusted with the number of iterations, which made the search step size of the improved algorithm limited and accelerated the convergence of the algorithm. Kapur entropy was selected as the adaptation function, and the improved crow algorithm was used to perform multi-threshold segmentation on the different types of images in the end, and the results of the algorithm in this article were compared with the segmentation results of the four algorithms such as the traditional crow algorithm, cuckoo algorithm. Lena, Flower, Fruits and Boat were segmented, and the structural similarity of improved crow search algorithm was 0.7703, 0.7761, 0.7276, and 0.7921; the standard deviations were 0.0295, 0.0385, 0.0344, and 0.0173. Experimental data shows that the improved algorithm had better segmentation than other algorithms. The algorithm in this paper effectively improves the blindness and the shortcomings of being easy to fall into the local optimum about the traditional crow algorithm. It can accurately segment complex images, which has certain reference values in the field of multi-threshold image segmentation.