Currently,　leukocyte　detection　has　the　problem　of　scarcity　of　labeled　samples,　so　a　focal　dataset　must　be　expanded　by　merging　multiple　datasets.　At　the　same　time,　given　the　difference　in　the　dyeing　methods,　dyeing　time,　and　collection　techniques,　some　datasets　have　the　problem　of　different　homology　distributions.　Moreover,　the　effect　of　direct　training　after　dataset　merging　is　not　satisfactory.　The　morphology　of　the　leukocyte　types　is　also　variable　and　stain　contamination　occurs,　thereby　leading　to　the　misjudgment　of　using　traditional　convolutional　networks.　Therefore,　in　this　paper,　the　model　parameter-transfer　method　is　used　to　alleviate　the　problem　of　less　leukocyte　labeled　data　in　the　training　model　and　deformable　convolution　is　introduced　into　the　main　network　of　target　detection　to　improve　the　accuracy　of　the　object　detection　model.　First,　numerous　leukocyte　datasets　are　used　to　train　the　blood　leukocyte　binary　classification　detection　network,　and　the　model　parameters　of　the　blood　leukocyte　binary　classification　detection　network　are　transferred　to　the　blood　leukocyte　multi　classification　detection　network　through　the　transfer　of　model　parameters.　This　method　can　make　better　use　of　datasets　of　the　same　origin　and　different　distributions　so　as　to　solve　the　problem　of　scarcity　in　blood　leukocyte　data　sets.　Finally,　the　multi　classification　detection　network　is　trained　quickly　and　the　accurate　blood　leukocyte　detection　results　are　obtained　through　fine　tuning.　The　experimental　results　show　that　compare　our　method　with　the　traditional　Faster　RCNN　object　detection　algorithm,　mAP0.5　is　0.056　higher,　mAP0.7　is　0.119 higher,　with　higher　recall　by　4%,　and　better　accuracy　by　5%.　Thus,　the　method　proposed　in　this　paper　can　achieve　highly　accurate　leukocyte　detection.　©　2021,　Springer　Nature　Switzerland　AG.