The rapid and accurate identification of citrus fruit is of great significance for the realization of automatically picking citrus in orchards，the prediction of citrus yield and the intelligent management of citrus industry. A citrus recognition method based on the improved model of YOLOv5 was proposed to realize the recognition of citrus fruits in natural environment.The feature extraction ability of the network and the problem of missed detection of occluded targets and small targets was improved by introducing the CBAM attention mechanism module. The α-IoU loss function instead of the GIoU loss function was used as the bounding box regression loss function to improve the positioning accuracy of the bounding box. The results showed that the average accuracy AP value of the proposed model reached 91.3%，with the detection time of a single citrus fruit image on the GPU of 16.7 ms and the model occupying 14.5 Mb of memory. It is indicated that the algorithm improved can quickly and accurately identify citrus fruits in the natural environment，meeting the practical application requirements of real-time target detection. It will provide new ideas for the intelligent citrus industry.