The invention relates to a leukocyte extraction and classification method based on an improved K-means and a convolutional neural network. The method comprises the steps of firstly, selecting an initial clustering center according to cell image gray level distribution, and clustering all pixels of an image initially according to the principle of proximity; then, improving the Euclidean distance ofthe FWSA-KM algorithm; before the extraction of leukocytes, carrying out the color space decomposition firstly, and carrying out the cell nucleus and cytoplasm extraction by adopting a color component beneficial to leukocyte segmentation and an improved K-means algorithm; separating a complex adhesion part by adopting a watershed algorithm; and finally, performing classification based on the convolutional neural network. According to the method, the leukocyte nucleus segmentation precision and the cytoplasm segmentation precision are 95.81% and 91.28% respectively, and compared with a traditional segmentation method, the precisions are greatly improved, the classification accuracy can reach 98.96% at most, the classification average time is 0.39 s, and compared with an existing leukocyteclassification algorithm, the CNN classification method not only has obvious advantages, but also has the great improvement space.