Honeysuckle flower stage identification based on improved YOLOv5s

Abstract

The medicinal constituents of Chinese herbal medicine honeysuckle (Lonicera japonica Thunb) vary at different flower stages. In order to ensure that the medicinal value is maximized, it is necessary to identify its flower stage before harvesting. However, at present, this study can only be accomplished by manual visual recognition, which is inefficient and costly. Therefore, there is an urgent need to develop an automatic detection technique with high maturity, fast detection speed, and strong model deployment capability. In order to adapt to the problems of different flower size and color texture similarity and complex background, this study chooses YOLOv5s algorithm for adaptive modification. First, a small detection layer is added to the network to enhance feature extraction and improve the accuracy of identifying small honeysuckle. Second, attention mechanism is incorporated into the backbone network to suppress background interference and improve identification accuracy. Finally, the original IoU-NMS is replaced by the DIoU-NMS algorithm, which improves the bounding box regression rate while reducing the leakage rate when overlapping or occluded. The test results showed that the P was increased from 80.0% to 92.7%, the R was increased from 78.6% to 80.2%, and the mean average precision was increased from 86.2% to 90.6%. Furthermore, the model was verified at both long range and short range, and the tests data indicate that the identification accuracy was no less than 90% in 3 m without serious occlusion. This study laid a solid foundation for accurate honeysuckle flower stage identification and provided technical support for real-time machine picking honeysuckle.