SP-RTSD: A Lightweight Real-Time Strawberry Detection on Edge Devices for Onboard Robotic Harvesting

On-board strawberry-picking robots offer the potential to significantly reduce labor costs and enhance picking efficiency. How to achieve high precision and fast strawberry recognition on resource-constrained edge devices is the key to robotic strawberry harvesting. Before developing our model, two lightweighting methods that maintain model structure are explored to substantiate the thesis that only judicious compression strategies tailored to edge hardware specifications can transform heavyweight deep models into efficient and compact deployments with enhanced performance on embedded devices. Based on this, and in combination with RTSD, Superb Real-time Strawberry Detection (SP-RTSD), which is designed to achieve faster and more accurate strawberry recognition on edge devices. Firstly, the C2f-Faster module performs channel-wise feature screening to enhance feature extraction efficiency while reducing model parameters; secondly, a lightweight recognition head with a parameter sharing mechanism is proposed specially for the edge devices. The speed of SP-RTSD was significantly improved by 22% from 20.63 to 25.18 FPS, which is similar to the 25.2 FPS of RTSD. Without changing the model structure, the model size is reduced by 40.3% from 6.2 to 3.7 MB. In contrast to typical lightweight strategies, which often boost inference speed at the cost of accuracy, SP-RTSD achieves exceptional accuracy with a mean average precision (mAP) of 91.7%, slightly outperforming the original baseline model (90.7%). The improvements in accuracy, speed, and size demonstrate that SP-RTSD addresses the challenge of balancing accuracy with inference speed on edge devices. In comparison experiments with other advanced object detection and lightweight models, as well as tests on additional open-source strawberry data sets, SP-RTSD consistently delivered superior results, affirming its robustness. Furthermore, SP-RTSD demonstrated an impressive combined success rate of 92% in strawberry grasping simulation experiments with a robotic arm, thereby confirming its suitability for integration into practical picking machines.