Steel Pipe Automatic Identification System Based On Image Processing

At present,most domestic bar producers use photovoltaic technology or manual methods to calculate the bar.The former is only applicable to the case where the bars do not overlap.And the latter counting process takes a long time,as well as the efficiency is low,and the labor intensity is also large,that is,cumbersome and error-prone.At the same time,many scholars have made efforts to make breakthroughs in algorithms to solve this problem,liberate productivity,save production costs,and reduce labor.However,in the field of bar counting,there is always no very good solution to solve such a problem.Either simply use manual counting.This method is labor intensive and has a high probability of error.Sometimes it needs to be repeated several times to count correctly,and a slight oversight will make mistakes,the cost of time and the cost of labor.It is very expensive,laborious and time consuming.Either a strict counting scheme will be designed to achieve accurate counting in a specific scenario,under certain conditions,plus image algorithm and manual review.In order to realize a convenient and fast bar counting scheme,this paper introduces an automatic steel pipe identification system based on target detection.Before establishing the algorithm,go to the actual site survey to understand the actual site bar stacking situation,then collect data on site and record video on several stacked steel pipe stacks.The data collected on the site site is pre-processed,and several collected video data are analyzed first,and after the data is cleaned and filtered,the bad data is cleared.Then the video data is converted into picture data frame by frame,and the data is randomly selected for trial processing.It is found that the traditional image processing method can be used,and the effective features cannot be extracted,and the edge of the steel pipe is extracted intact,then the next count cannot be Go on.Then try other ways to identify.In the attempt to use deep learning to achieve the purpose of this article,first pre-process some of the data,including data cleaning,data enhancement,data annotation and so on.And use the structure of yolov3 in deep learning to train,in order to get a good model.Then extract a part of the remaining data to test the model,and found that the result is good,decided to use YOLOv3 to solve the problem.This article uses NVIDIA1070 TI graphics card for GPU floating point calculation,using YOLOv3 training model,then transplant the model,add pyqt program design,and improve to make a system.The test results are good and easy to use.