Human peripheral blood leukocyte five kinds classification plays a vit al role in clinical diagnose.In this paper,we propose a method of human peripheral blood leukocyte microscopic image classification based on deep Convolutional Neural Network (CNN).Firstly,a deep CNN architecture based on ResNet is designed and it is applicable to leukocyte microscopic image.Then a new data augmentation metho d based on feature focus is proposed to enrich our dataset.Considering that the surroundings are important for object recognition,we generate a large number of images by putting a segmented leukocyte in different microscop ic image surroundings using image processing.After data augmentation,the total amount of the dataset is 42300.Finally,aiming at the disproportion of five kinds of leukocyte in dataset,we propose a modified batch stochastic gradient descent (MBGD) to train the CNN model.By setting the ratio of five kinds of leukocytes into 1∶1∶1∶1∶1in a batch,CNN model can evenly achieve features of five kinds of leukocyte.Experimental resul ts demonstrate that the designed CNN architecture,proposed data augmentation method and modified stochastic grad ient descent can all improve the classification accuracy.The proposed method can achieve 95.7% training accu racy.The average testing accuracy of 8400leukocyte images is 95.0%.The accuracy of ne utrophils,lymphocytes,monocytes,eosinophils and basophils are respectively 92.2%,91.5%,94.6%,93.3% and 97.4%.