Research On Lightweight Parts Recognition Algorithm For Augmented Reality Applications

At present,in the field of industrial assembly,the main assembly instruction method of workers is still based on referring to the manual to guide the assembly process.This type of instruction usually faces problems such as low assembly efficiency and high requirements for the experience of the assembly personnel.The augmented reality technology benefits from its good experience and interactive technology,and it is widely used in entertainment,education,and industry.However,the recognition and detection algorithms involved in the current augmented reality technology are usually detection algorithms based on image texture features.This type of algorithm usually requires the object to be detected to have high contrast and complex texture patterns.For industrial parts with few features and single texture,the recognition effect is usually not ideal.Therefore,it is necessary to study a part detection algorithm suitable for augmented reality applications.In the field of target detection,excellent detection networks such as Faster R-CNN,SSD,YOLO,etc.are widely used in various fields,but such networks are usually accompanied by a large number of parameters,resulting in a model that is too large.The hardware performance requirements are relatively high,which is not convenient for deployment on resource-constrained augmented reality devices.Secondly,since the training of the target detection network usually requires a large number of data set samples for training,in the actual industrial production process,the labeling of the data set usually brings a large time cost,so the data labeling problem has become an urgent problem to be solved.In response to the above problems,this article has done the following work:(1)In terms of lightweight network model construction,this paper optimizes the backbone network structure in YOLOv3 to compress the network model,and at the same time improves the feature fusion structure in the original YOLOv3 to improve the detection accuracy of the network.(2)In terms of data set production,this article uses three-dimensional virtual data to construct part data samples,which greatly reduces the cost of labeling time.At the same time,the size,pose and lighting conditions of the detected parts can be controlled by scripts.Has a good data enhancement effect.Compared with the original target detection network,the optimized YOLOv3 target detection network in this paper has less model accuracy degradation,and the parameter quantity of the model has been greatly compressed.The size of the optimized network model in this article is only 4.3M,in terms of detection accuracy,the optimized network has reached 97.57% m AP of the parts data set in this article.Through testing on mobile devices,the part detection network designed in this paper can detect a single image at a speed of 124 ms.In summary,the optimized YOLOv3 network in this article meets the actual functional requirements in terms of model size and reasoning accuracy,and is conducive to the deployment of models on resource-constrained devices,and has certain practicability.