The number and proportion of five kinds of leukocyte in human peripheral blood reflect the health state of human body. Manual examination of leukocyte consumes a lot of manpower for medical?workers. How to use intelligent method to classify leukocytes quickly and accurately is an urgent problem to be solved. The accuracy of leukocyte segmentation is the key to correct classification. In this paper, an improved iterative threshold image segmentation algorithm is proposed, and the minimum distance method for restoring mitotic lines is improved based on mathematical and digital simulation analysis. The accuracy and efficiency of leukocyte segmentation are improved, and the problems of platelet adhesion and unclear leukocyte boundary are solved by these methods. The leukocytes are separated from the complex blood environment, the minimum distance of the lobulated nucleus leukocytes is determined and connected, and then located to each leukocyte to make a data set, finally, it is classified by CNN. After testing, the accuracy of leukocyte segmentation is more than 96%. The experimental results show that the proposed segmentation method is accurate, efficient and practical.