The statistics of pollination times of honeybee was evaluated based on machine vision to evaluate the rationality of the dosage of pollinating honeybee during the window period of blueberry blooming in the solar greenhouse. The dataset was processed with the method of improved poisson blending data enhancement to solve problems that the detection environment is complex， the target scale is small， and it is easy to be covered. The structure of YOLOv5 was optimized. The detection precision of model was improved by introducing GAM attention mechanism and Transformer module. BiFPN and CARAFE were introduced in feature pyramid network to complement the contextual information. EIoU_loss and Soft NMS were used to enhance the positioning precision of bounding box and solve the problem of detecting target occlusion. The results showed that the mean average precision of the improved YOLOv5 was 96.6%， 3.5 percentage points higher than that of the original algorithm. The detection time of a single blueberry pollination image on the GPU was 11.4 ms.