In order to solve the problem of the adhesion segmentation in fluorescent microspheres image and the classification with limited labeled samples, we propose a fluorescent microsphere segmentation and classification method based on improved watershed and semi-supervised minor reconstruction error classifier (SSMREC). Firstly, we use improved watershed method to segment the fluorescent microspheres adhesion image, and effectively separate the adhesion fluorescent microspheres into independent objects. Then we use the non-uniform quantization of Hue-Saturation-Value (HSV) color space for the microsphere objects to remove the redundant information and extract the discriminant features. Finally, the microsphere objects are classified by a semi-supervised reconstruction error classifier. We compare the proposed method with linear discriminant analysis classifier (LDA), random forest classifier (RFC), sparse representation-based classifier (SRC), K- nearest neighbor classifier (KNN), and support vector machine (SVM). The results show that the overall classification accuracy of the proposed method can be improved by 3.5%-14.3% compared with other classification methods in the case of randomly selecting 2, 4, 6, and 8 labeled samples in each class. It means that the proposed method is more effective in small number of labeled samples.