In order to improve the accuracy of cherry fruit recognition and the working efficiency of orchard automatic picking robot, totally 1816 sets of cherry original images collected in Yantai Academy of Agricultural Sciences and data images obtained with different data enhancement methods were used to establish the data set, the data set was divided into training set and test set according to rate of 8∶2, and YOLO v5 model was used to identify and detect cherry data sets enhanced by different data enhancement methods and combined enhancement methods based on the in-depth learning network. The results showed that offline enhancement and online enhancement had a certain positive effect on the improvement of model accuracy. The offline data enhancement strategy could significantly and stably increase the detection accuracy, and the online data enhancement strategy could slightly improve the detection accuracy. Using the combination of offline enhancement and online enhancement at the same time could greatly improve the average detection accuracy. In view of the mutual occlusion between cherry fruits and the difficulty in detecting small cherry targets in the picture, the detection accuracy of cherry fruits in the natural environment was low, the backbone network of YOLO v5 was changed, the transformer module with attention mechanism was added, and the neck structure was changed from the original pafpn to bifpn which could carry out two-way weighted fusion. The P2 module of shallow down sampling was added to the head structure. The experimental results showed that compared with other existing models and the improved YOLO v5 model with a single structure, the comprehensive improved model proposed had the highest detection accuracy, and the mAP@0.5∶0.95 was increased by 2.9 percentage points. The results showed that this method effectively enhanced the efficiency and accuracy of feature fusion in the recognition process, and significantly improved the detection effect of cherry fruit. At the same time, the trained network model was deployed on the Android platform. The system was simple and clear to use, and the requirements of user equipment environment were not high. Therefore, the system had certain practicability. It could detect cherry fruit in real time and accurately in the field environment, which laid a foundation for practical applications such as automatic service picking in the future.