RLK-YOLOv8: Multi-Stage Detection of Strawberry Fruits throughout the Full Growth Cycle in Greenhouses Based on Large Kernel Convolutions and Improved YOLOv8

Lei He ^1,2,3 Dasheng Wu ^1,2,3* Xinyu Zheng ^1,2,3* Fengya Xu ^1,2,3 Shangqin Lin ^1,2,3 Wang Siyang ^1,2,3 Fuchuan Ni ⁴ Fang Zheng ⁴

• ¹ College of Mathematics and Computer Science, Zhejiang Agriculture and Forestry University, Hangzhou, China
• ² Key Laboratory of State Forestry Administration on Forestry Sensing Technology and Intelligent Equipment, Hangzhou, Jiangsu Province, China
• ³ Key Laboratory of Forestry Intelligent Monitoring and Information Technology of Zhejiang Province, Hangzhou, Jiangsu Province, China
• ⁴ Engineering Research Center of Intelligent Technology for Agriculture, Ministry of Education, Wuhan, China

In the context of intelligent strawberry cultivation, achieving multi-stage detection and yield estimation for strawberry fruits throughout their full growth cycle is essential for advancing intelligent management of greenhouse strawberries. Addressing the high rates of missed and false detections in existing object detection algorithms under complex backgrounds and dense multi-target scenarios, this paper proposes an improved multi-stage detection algorithm RLK-YOLOv8 for greenhouse strawberries. The proposed algorithm, an enhancement of YOLOv8, leverages the benefits of large kernel convolutions alongside a multi-stage detection approach.Method: RLK-YOLOv8 incorporates several improvements based on the original YOLOv8 model. Firstly, it utilizes the large kernel convolution network RepLKNet as the backbone to enhance the extraction of features from targets and complex backgrounds. Secondly, RepNCSPELAN4 is introduced as the neck network to achieve bidirectional multi-scale feature fusion, thereby improving detection capability in dense target scenarios. DynamicHead is also employed to dynamically adjust the weight distribution in target detection, further enhancing the model's accuracy in recognizing strawberries at different growth stages. Finally, PolyLoss is adopted as the loss function, which effectively improve the localization accuracy of bounding boxes and accelerating model convergence.The experimental results indicate that RLK-YOLOv8 achieved a mAP of 95.4% in the strawberry full growth cycle detection task, with a precision and F1-score of 95.4% and 0.903, respectively. Compared to the baseline YOLOv8, the proposed algorithm demonstrates a 3.3% improvement in detection accuracy under complex backgrounds and dense multi-target scenarios.Discussion: The RLK-YOLOv8 exhibits outstanding performance in strawberry multi-stage detection and yield estimation tasks, validating the effectiveness of integrating large kernel convolutions and multi-scale feature fusion strategies. The proposed algorithm has demonstrated significant improvements in detection performance across various environments and scenarios.