关键词: 苹果, 多目标识别, 采摘机器人, 神经网络

Abstract: Aiming at the issues of multiple influencing factors and low recognition accuracy in multi target detection of harvesting robots, an improved YOLOv5 object recognition network model is proposed to provide effective visual guidance for harvesting robots. Firstly, based on YOLOv5s, an improved Ghost module is used to replace the CSP module to reduce the computational complexity of the model and make it more lightweight. Considering the aspect ratio between the prediction box and the target box, CIoU_Loss is replaced by GIoU_Loss as a loss function to improve the stability of target regression. DSPP module is used to replace the original SPP module for pooling operation to solve the problem of SPP feature information loss and improve the accuracy of object detection. Then, the model is used to train and evaluate apple image samples,to analyze the reliability and feasibility of the models performance. Finally, identification testing is conducted on the dataset and comparative analysis is conducted based on different evaluation indicators to verify the superiority of the constructed model. The research results indicate that the proposed improved YOLOv5 network model has the characteristics of high recognition accuracy, strong antiinterference ability, and fast detection speed. Compared with the other models, the accuracy rate is up to 8.4%, the recall rate is up to 5.1%, and the mAP rate is up to 10.3%.

Key words: apple, multitarget recognition, picking robot, neural network