Research On Industrial Target Detection Algorithm Based On Edge Computing

Research on target detection algorithm in industrial scene is the focus of the current manufacturing industry to automation and intelligent upgrade process.This research uses computer vision technology as a means to find target objects in pictures or video data to complete tasks such as product quality inspection,part identification and operator code of conduct inspection.Is different from other scene target detection task of algorithm precision acme requirements,because of its special scenario requires industrial scene,need detection algorithm has a high real time capability,maintain a balance between accuracy and speed of requirements algorithm,especially the various considerations for cost,power consumption,etc.,the industrial scenario target detection algorithm are often deployed in computing ability is on the bottom of the embedded devices,The edge of calculation,the calculation of the algorithm itself put forward higher requirements,request algorithm can not only win by "quantity",resource usage is also examining algorithm performance in industrial situations one of the main indicators,even need algorithm in terms of accuracy to make certain concessions in exchange for a higher detection rate,guarantee real-time industrial detection.In order to better meet the target detection task based on edge devices in industrial scenarios,this paper improves the original Yolov5 algorithm by using lightweight convolution components Ghost Conv and Ghost Bottleneck,which reduces the computation resource consumption of the algorithm and improves detection efficiency.In addition,a lightweight attention mechanism Coord Attention is introduced to improve the lightweight optimization of the model without bringing extra load to the computing device,leading to a loss of precision.Combined with structured pruning technology,the computation of the algorithm is further reduced.The size of the new improved model is only 37.2% of that of the original Yolov5,and the calculation time is reduced by 40.9%.To fully restore the target detection algorithm in the application of industrial scenario,this paper will be deployed into the model to the edge of the side is one of the research task will be trained in the GPU environment detection model transformation,generation can be conducted on embedded devices offline model of reasoning algorithms,the offline model on the precision and speed can meet target detection in the actual industry mission requirements,In addition,in order to enhance the detection effect of the model,this paper makes targeted improvements to the data set used,and enhances the data with image processing technology to improve the detection accuracy and generalization ability of the model.