Research On DELTA Robot Vision Recognition System For Food Grasping

With the continuous development of robotics and artificial intelligence technology,robots based on vision technology have been widely used in positions with high repetition,such as food sorting.However,the traditional vision technology has poor anti-interference ability and cannot adapt to complex production environment.To solve this problem,a DELTA robot visual recognition system is designed in this thesis.To balance detection speed and accuracy,a lightweight model,YOLOv5-MGCS,is proposed to realize food detection.The main contents of this thesis are as follows:(1)To build the food dataset,we first collected 2049 food images,then used data enhancement operations on them,and then annotated the expanded images to finally complete the food dataset.(2)The YOLOv5-GCS model was constructed on the basis of YOLOv5 s.CBAM attention mechanism is introduced into the feature fusion network of YOLOv5 s to strengthen the model’s attention to detection targets,suppress background information,and improve the antiinterference ability of the network.Ghost convolution is used to replace the common convolution of feature fusion layer,Ghost convolution can reduce the generation of redundant feature graphs,reduce the number of model parameters,so as to reduce the cost of calculation.To solve the problem of low positioning accuracy of YOLOv5 s,SIOU Loss was used as position loss function to improve the positioning ability of the model.The results show that the detection performance and positioning accuracy of YOLOv5-GCS are improved.The comparison experiment with multiple single-stage target detection algorithms shows that YOLOv5-GCS has the strongest comprehensive performance and good stability.(3)The lightweight detection algorithm YOLOv5-MGCS is proposed after the lightweight processing of YOLOv5-GCS.The first 17 layers of Mobile Netv3-large were selected to replace the feature extraction network of YOLOv5-GCS.Compared with YOLOv5-GCS,the number of model parameters after lightweight is greatly reduced,which can ensure the detection accuracy while maintaining a higher detection speed.(4)A food vision recognition system was designed based on YOLOv5-MGCS.The overall architecture and functional modules of the system were designed first,and then the software of the detection system was developed using Py Qt and Python.Finally,the software is tested based on the YOLOv5-MGCS detection model.The test results show that the detection software can achieve the functions of file selection,model weight selection,image display,etc.,which verifies the reliability of the detection system.