The work aims to propose an improved K-means high-throughput dPCR fluorescent image classification algorithm to achieve high-precision classification of high-throughput dPCR fluorescent image positive points. Firstly, the gray value of the preprocessed fluorescent image was counted, and wave peak and valley adaptively were selected according to image brightness to determine cluster center. The pixel clusters were determined by Mahalanobis distance. Then, the broad classification results were processed by morphology opening-and-closing operations and deleting small area objects. Finally, the fine classification, location identification and counting were completed with the third connected domain statistics. In the experiment, 825 fluorescence images of four channels were selected to test. The average accuracy was 99.06%, and recall rate was 98.97%, showing good classification effect. The classification algorithm of improved K-means proposed in this paper can achieve high-precision classification and counting of high-throughput dPCR fluorescent images, and can be used for reference to other fluorescent image classification and recognition.