Research On Target Recognition Algorithm Of Apple Picking Robot Based On Vision

As the largest output of the main export fruit,apple picking is still completed by manual,due to the continuous improvement of labor costs,it is an inevitable trend to use robots to pick apple.The identification and detection of apples in the process of apple picking is the basis and guarantee of the whole work of apple picking.Therefore,in order to ensure the stable and efficient work of apple picking,the key technologies involved in the work of apple picking robot are studied by deep learning algorithm combined with the actual situation.The research content includes four aspects:(1)Establishment and expansion of data sets.In order to solve the problem of lack of existing unpicked apple data set,this paper obtains part of unpicked apple data set through web crawler and actual shooting.In order to meet the amount of data required for model training,this paper expands the data set through translation,rotation,changing brightness and contrast,adding noise and other operations,and then marks according to the labels required for network training.(2)Apples were not picked for detection.According to the actual load requirements of apple picking robot,lightweight optimization of YOLOv4 network model structure was carried out.Firstly,Ghost Net network structure was replaced by YOLOv4 backbone network,and Fire Module and Bi FPN network structure were used to optimize the structure of YOLOv4 feature fusion module.Then,the feasibility of each improvement point to optimize the network model was proved by ablation experiment.The operation of modifying the backbone network and the traditional convolution reduced the model size by 82.59% on the premise that the model accuracy was almost unchanged.The model accuracy was improved by 2.29% on the premise that the model weight file size was reduced by 86.31%,and the model detection efficiency was improved by 23.95%.Through the experiment on VOC data set,the m AP value is 74.85%,which verifies that the model still has good generalization and robustness.(3)Apple segmentation and picking point determination.The traditional method of extracting picking points is complicated and requires a lot of artificial parameters,so the result is not satisfactory.In this paper,the classical algorithm YOLACT in case segmentation was used to segment the unpicked apple contour from the background,and the apple picking point was extracted according to its contour.Through experiments,the average accuracy of the model in the prediction frame reaches 98.9%,and the average recall rate reaches 94.4%.The average accuracy of the mask reached96.6%,and the average recall rate reached 85.8%.And through the actual detection results confirm that the model has a good effect in unpicked apple segmentation and picking point determination.(4)Apple grading.In order to improve the efficiency of apple picking,apples were effectively graded after picking.The Apple-data sets with different data-enhanced lesions are tested using the Efficientdet algorithm,which improves its training parameters.The results showed that the improved model could effectively distinguish normal apples,apples with skin damage and apples with pulp damage.And the accuracy of classification results of all kinds of damage meets the actual demand,which can be applied to the rapid classification of apples after actual picking.