A Study On Model Rapid Training And Frame Rate Optimization For AR Aided Assemble System

This thesis comes from the crosswise tasks of locomotive bolt AR auxiliary assembly project,which is oriented to locomotive overhaul process and proposes a bolt auxiliary assembly scheme based on AR,Internet of Things and deep learning technology.The image processing device in the auxiliary assembly system is a target detection and recognition device based on deep learning.Under the Internet of Things technology,it can communicate with other intelligent devices in the auxiliary assembly system.The main research content of this thesis is about the rapid model training and frame rate optimization technology of the deep learning platform.In the process of using the auxiliary assembly system,if the accuracy of the model is low,it is necessary to collect the image again and retrain the model,and as a new locomotive models to join the production,also need to train corresponding model of new models,and the model of training time usually lasted for several days,to carry out the training of the model is of a very consumption time task.Therefore,it is of great significance to speed up model training for auxiliary assembly system.In the assembly process,the image processing link will block the subsequent assembly process.Therefore,the lower the server's image processing speed is,the slower the server's response speed will be,and the less smooth the assembly process will be.The higher the frame rate is,the faster the server's response will be,and the smoother the assembly process will be.Therefore,improving the frame rate is a key to improve the performance of auxiliary assembly system.The main contents and corresponding results discussed in this thesis are as follows:First,the network structure of the target detection algorithm YOLOv3 was optimized,so that the improved network could shorten the training time of the model on the premise that the loss index did not decrease,and the transfer learning technology was applied to reduce the overfitting phenomenon caused by the training model of small data set,and further shorten the training time.Secondly,the number of parameters and computation of the network are optimized by separable convolution,and the computational speed and frame rate of the model are improved.Thirdly,in order to further improve the frame rate of image recognition,a fast image transmission channel based on Web Socket is implemented to improve the frame rate of image recognition by improving the transmission speed between the image acquisition equipment and the server.Fourth,in order to facilitate the management and development of training networks with different parameters,a deep learning platform for locomotive bolt assembly scene is designed and realized,which supports the free customization of important network parameters.