A mulberry detection method in the complex background based on improved D-YOLOv3

Currently, mulberry has important crop value. In the densely planted field, many mulberry trees would be overlapped and occluded to varying degrees. To solve the mulberry tree identification problem based on overlapping and occlusion environment, the improved YOLOv3 model was proposed in this study. The improved method uses DensNet's dense connection mechanism to replace the last three sub-sampling layers of feature extraction network DarkNet53 in the YOLOv3 network to reuse features and retain important semantic features. Secondly, use the CloU loss instead of the IoU loss in YOLOv3 to improve the speed of target recognition and achieve better detection performance. And the K-means clustering is used to set the parameters of the mulberry target anchor box, which improves the accuracy of complex target recognition and location. Moreover, a transfer learning strategy was adopted to help train new models, which can greatly enhance the generalization ability of the model. The experiment tests the improved and original methods' detection performance. And the influence of different training methods on the model, respectively. The results showed that the proposed model could effetely detect mulberry trees based on overlapping and occlusion environment. It is of great significance to provide the theoretical vision basis for intelligent mulberry leaf picking in the future.