The invention discloses an automatic stem cell counting method based on depth learning, which relates to the technical field of cell counting. The method comprises the following steps: S11, a phase contrast microscope is adopted to photograph cells in need of counting, and a stem cell image is generated; S12, image preprocessing is carried out, and the cell image is subjected to noise reduction processing and illumination equalization processing to obtain a stem cell image with equalized illumination; S13, a cell artifact generated during the photographing process by the phase contrast microscope is removed; S14, the stem cell image after the artifact is removed is segmented, and multiple candidate stem cell images are acquired; S21, the segmented multiple candidate stem cell images are manually marked, and a training set is built; S22, the training set is inputted to a CNN for training; and S23, a stem cell counting result is counted. The defect that a large amount of manpower is consumed in the traditional manual cell counting can be solved, the defect that a flow counting method damages the cell growth environment is overcome, counting is carried out through the photographed cell image, and the method has the advantages of being stable, efficient, automatic and lossless.