MS-SFNet: a tiny reproductive phenological stage detection model for strawberry inspection robots

Accurate phenological perception is a critical prerequisite for intelligent lifecycle management of facility strawberries. However, current field inspection practices are constrained by deployment difficulties and insufficient real-time performance. To address these issues, this paper proposes MS-SFNet, a lightweight phenological detection model tailored for inspection robots. A three-scale detection architecture is constructed to handle the pronounced scale variability across strawberry phenological stages, incorporating a layered spatial-semantic fusion module and a multi-kernel residual perception module to enhance multi-scale feature representation while substantially reducing model size. Depthwise separable convolutions are adopted in the feature aggregation pathway to expand the receptive field while mitigating the loss of small-target features caused by downsampling. A CB-FC-NW compound loss function is devised, introducing independent class-specific scaling coefficients for rare categories to alleviate the long-tail distribution problem in real-world strawberry field datasets. Training is conducted on an NVIDIA RTX 4090 GPU. Experimental results show that the model achieves only 0.89 M parameters and an inference speed of 278.0 FPS. Compared with over a dozen mainstream detectors, our model attains optimal overall performance. It delivers superior detection accuracy for four phenological stages directly tied to yield estimation and harvesting: Ripening (0.947), Expansion (0.889), Color-Change (0.884), and Flower (0.842). With the proposed loss function, the mAP50 for the most challenging Bud category improves from 0.448 to 0.488, a gain of 4 percentage points, effectively resolving the trade-off dilemma of conventional loss functions on long-tail classes. MS-SFNet achieves a favorable balance among extreme lightweightness, real-time responsiveness, and long-tail suppression, demonstrating strong deployment potential on self-developed strawberry inspection robots.