Abstract: Aiming at the problems of overlapping fruits， interference of branches and leaves， and complex backgrounds in apple orchards， the Faster-RCNN algorithm was proposed. By adding Mosaic data enhancement at the input end， the amount of data is increased and the ability to recognize small objects is enhanced. At the same time， the anchor frame in the Faster-RCNN algorithm is optimized， and the optimized anchor frame can better detect the distance. The target fruit far from the camera and the Soft NMS algorithm are used to further improve the recognition effect of dense areas. The verification results show that the recall rate is 91.44%， the accuracy rate is 93.35%， the F1 value is 92.38%， and the average detection speed per image can reach 0.2 fps. The robustness of the improved algorithm is enhanced， which can meet the recognition of apple fruits in natural environment.

Key words: Faster-RCNN, Mosaic data augmentation, target recognition, Soft NMS algorithm