Fruit Target Detection Technology In Orchard Environment Based On YOLOv4

With the continuous promotion of smart agriculture,fruit picking robots based on computer vision have become a new development trend.The key to automatic fruit picking lies in the accurate detection of fruit targets,but the accuracy of fruit detection in the orchard environment is often affected by factors such as overlapping and occlusion,resulting in problems such as omissions in picking.Therefore,this study takes the fruit overlapping occlusion problem in the orchard environment as the research background and takes apple,orange,strawberry,watermelon and springsee citrus as the research objects,aiming to improve the detection accuracy of overlapping occlusion fruits in the orchard environment.The specific work content is as follows:First of all,for the situation that there is no public fruit dataset in the orchard environment,which cannot meet the experimental requirements in the complex orchard environment,this thesis builds a common fruit dataset in the orchard environment.At the same time,aiming at the problem that the traditional data enhancement method is insufficient to improve the localization ability of the model,a data enhancement method of branch and leaf illustration is proposed to expand the original image dataset and improve the robustness of the network model.Secondly,aiming at the situation that the detection accuracy is insufficient due to the overlapping occlusion between the fruits to be detected or the occlusion by branches and leaves in the real orchard environment,a fruit target detection algorithm based on the overlapping occlusion is proposed.On the basis of the YOLOv4 network,the MobileNetV3 network is introduced as the backbone feature extraction network,the K-means++ algorithm is used to optimize the parameters of anchor,and the bidirectional feature pyramid networks(Bi-FPN)and the improved spatial coordinate attention(SCA)integrate contextual information to improve the localization and detection capabilities of the model.Experiments show that the average accuracy of the model algorithm reaches 93.02%,which can achieve accurate detection of overlapping occluded fruits in orchard environment to a certain extent.Finally,for the automatic picking of bagged fruits on the mobile terminal,a lightweight bagged fruit detection algorithm is proposed.In this thesis,the lightweight YOLOv4-tiny network is selected as the backbone feature extraction network,and spatial pyramid pooling structure(SPP),the improved spatial coordinate attention(SCA)and feature pyramid structure(FPN)are introduced into the network at the same time,and the YOLO Head detection layer of the YOLOv4-tiny model is replaced with the Center Head without anchor.In the detection experiment of bagged fruit,the accuracy of the network model reaches the highest average accuracy of 94.40%,and the model size is only 27.49 MB.At the same time,in the bagging fruit stalk detection experiment,the average accuracy of the model reaches87.97%,and the average detection time of each image is only 11 ms,which can meet the needs of real-time detection of mobile terminals.