Abstract: The intelligent grading and recognition of tea buds in a natural environment are fundamental for the automation of premium tea harvesting. To address the problems of low recognition accuracy and limited robustness caused by complex environmental factors like lighting, obstruction, and dense foliage, we propose an enhanced model based on YOLOv7-tiny. Firstly, a CBAM module was added into the small object detection layer of the YOLOv7-tiny model to enhance the model's ability to focus on small object features and reduce the interference of complex environments on tea bud recognition. We adjusted the spatial pyramid pooling structure to lower computational costs and improve detection speed. Additionally, we utilized a loss function combining IoU and NWD to further enhance the model's robustness in small object detection by addressing the sensitivity of the IoU mechanism to position deviations. Experimental results demonstrate that the proposed model achieves a detection accuracy of 91.15%, a recall rate of 88.54%, and a mean average precision of 92.66%. The model's size is 12.4 MB. Compared to the original model, this represents an improvement of 2.83%, 2.00%, and 1.47% in accuracy, recall rate, and mean average precision, respectively, with a significant increase of 0.1 MB in model size. Comparative experiments with different models show that our model exhibits fewer false negatives and false positives in multiple scenarios, along with higher confidence scores. The improved model can be applied to the bud grading and recognition process of premium tea harvesting robots.

Key words: YOLOv7-tiny, tea bud, grading identification, attention mechanisms, NWD loss