The deep learning technology have been applied in different domain successfully, it also has great potential in various agricultural applications. Real-time crop detection can provide a lot of information to agricultural robot and achieving various ta...
The deep learning technology have been applied in different domain successfully, it also has great potential in various agricultural applications. Real-time crop detection can provide a lot of information to agricultural robot and achieving various tasks. However the onions are small and in dense distribution, traditional detection model cannot perform well in this situation. In this research, an improved YOLOv3 model was proposed for dense onions detection. Onion images were collected at the conditions of different illuminations, different onions distributions, and occluded samples. The YOLOv3 model structure was simplified by truncating redundancy branches and layers. Then, the generalized intersection over union(GIoU) was introduced into loss function and so weights could be updated even no intersection between prediction and ground truth. The imbalance of inliers and outliers was reduced by applying focal loss method. The test result show that the proposed model was faster and has better generalization ability than original YOLOv3 model. The counting accuracy was over 96% in different onions distribution and the detection speed was improved by 30% than original method.