KM-Mask RCNN: A Lightweight Instance Segmentation Algorithm for Strawberries With Multiple Growth Cycles

Accurate and efficient detection of multi-growth-cycle strawberry fruits can improve automated harvesting. However, the small size and unbalanced distribution of strawberry fruits make accurate identification of multi-growth-cycle strawberries difficult using the existing detection models. Herein, local enhancement technology is adopted for preprocessing when compiling the dataset to ensure the balance of the number of samples in each category to solve the above problems. Second, a new instance segmentation algorithm called KM-Mask RCNN is developed, which optimally adjusts the size of the anchor frame and the anchor ratio based on the K-Means clustering algorithm to improve the recognition accuracy of the algorithm on small targets and uses MobileNet V3 to replace the Resnet50 structure in the Mask RCNN backbone network to reduce the complexity of the algorithm and realize lightweight operation. Finally, the experimental results reveal five strawberry growth stages in the homemade dataset(based on StrawDI_Db1 database): ‘Green ripe stage’, ‘White ripe stage’, ‘Turning stage’, ‘Mature’, and ‘Deformed’, for which the KM-Mask RCNN yields mAPs of 91.19%, 88.09%, 93.70%, 93.19%, and 87.13%, respectively. The P, R, and F1-score values of this algorithm are 93.9%, 94.2%, and 94.05%, respectively. Additionally, the number of parameters, FLOPs, and fps of this algorithm are 27M, 12G, and 22.32, respectively, satisfying the real-time requirements for strawberry detection. The findings provide important theoretical support for the automated harvesting of strawberries with multiple growth cycles.