Abstract:Aiming at the problem of camera shake and relative motion of objects leading to blurred detection images during target detection in orchards, a D2-YOLO one-stage deblurring recognition deep network that combined the DeblurGAN-v2 deblurring network and the YOLOv5s target detection network was proposed. It was used to detect and identify obstacles in orchard blurred scene images. To reduce the number of parameters of the fusion model and improve the detection speed, firstly the standard convolution used in the YOLOv5s backbone network with a deep separable convolution was replaced, then CIoU_Loss was used as the bounding box regression loss function of prediction. The fusion network used the improved CSPDarknet as the backbone for feature extraction. After recovering the original natural information of the blurred image, it combined multi-scale features for model prediction. To verify the effectiveness of the proposed method, seven common obstacles in the real orchard settings were selected as the target detection objects, based on the chassis of the crawler mobile robot, the BUNKER was equipped with portable computers, cameras and other equipment to form a mobile platform for image acquisition, and the model training and testing were carried out on the Pytorch deep learning framework. The precision and recall rates of the proposed D2-YOLO deblurring detection network were 91.33% and 89.12%, respectively, which were 1.36 percentage points and 2.7 percentage points higher than that of the step-by-step training DeblurGAN-v2+YOLOv5s. Compared with YOLOv5s, there was an increase of 9.54 percentage points and 9.99 percentage points in precision and recall rates, which can meet the accuracy and realtime requirements of orchard robot obstacle deblurring recognition. The research result can provide a reference for obstacle detection of agricultural robots in orchard in the later stage.
