Research on tea buds detection based on optimized YOLOv5s

As one of the world's most popular beverages, tea plays a significant role in improving tea production efficiency and quality through the identification of tea shoots during the tea manufacturing process. However, due to the complex morphology, small size, and susceptibility to factors like lighting and obstruction, traditional identification methods suffer from low accuracy and efficiency. In this study, image enhancement techniques such as HSV transformation, horizontal flipping, and vertical flipping were applied to the training dataset to improve model robustness and enhance generalization across varying lighting and angles. To address these challenges in the context of tea buds detection, deep-learning-based object detection methods have emerged as promising solutions. Nevertheless, current object detection technologies still face limitations when detecting tea buds under these conditions. To enhance identification performance, this article proposed an improved YOLOv5s (You Only Look Once version 5 small model) algorithm. In the improved YOLOv5s algorithm, CBAM, SE, and CA attention mechanisms were incorporated into the backbone network to augment feature extraction, and a weighted Bidirectional Feature Pyramid Network (BiFPN) is employed in the neck network to boost performance, resulting in the YOLOv5s_teabuds model. Experimental results indicated that the improved model significantly outperformed the original in terms of precision, recall, mAP and F1-score, with the CA attention mechanism providing the most notable improvement—enhancing precision, recall, mAP and F1-score by 18.119%, 9.633%, 16.496% and 13.524%, respectively. After integrating BiFPN, the YOLOv5s_teabuds model further strengthened performance and robustness, with precision, recall, mAP and F1-score increased by 19.346%, 11.388%, 18.620%, and 15.059%, respectively. Experimental results prove that the optimized YOLOv5s model can provide a real-time, high-precision tea buds detection method for robotic harvesting.