Research On Target Recognition Based On Machine Vision For Three-dimensional Imaging And Convolutional Neural Network

Safety is the primary task in industrial production,but the increasingly complex industrial site environment has brought great challenges to safe production.Commonly used safety monitoring methods in current industrial sites: 3D imaging methods based on machine vision,and intelligent recognition methods based on deep learning.Manual inspections are limited by human inspection accuracy,frequency and human reach.Therefore,industrial field inspection methods based on machine vision and based on deep learning have been widely used.These methods can be used in extremely harsh environments,but the degree of automation of the three-dimensional inspection method based on machine vision is still not high,and it often needs to be supplemented by human monitoring.In response to the above problems,firstly,this paper studies the 3D imaging method based on machine vision,and proposes a weighted calibration method based on Zhang's,and then this method is applied to the phase-shift fringe 3D imaging system using half-period reverse phase error compensation,which improves the measurement accuracy of the 3D workpiece size and imaging detection accuracy.Furthermore,this paper studies the detection method based on deep learning,and establishs a model of VGG16 combined with the early stopping method,which reduces the training cost and realizes the purpose of rapid detection.The main research work and innovations of this paper are as follows:1.This paper analyzes the distribution law of reprojection error of the calibration plate corner points,and proposes a weighted calibration method based on Zhang's.This method establishes a weighted objective function based on a two-dimensional Gaussian mathematical model,and obtains the optimal solution of camera parameters through iterative calculations,which improves the overall calibration accuracy of the system,and the calibration reprojection error is reduced from 0.0434 to 0.0232.The weighted calibration method is applied to the phase shift fringe projection 3D imaging system using half-period reverse phase error compensation.The system eliminates the periodic phase harmonic error caused by the nonlinear response,and effectively improves the 3D imaging display effect and the 3D measurement accuracy.2.This paper analyzes the noise model,combined with image filtering and segmentation algorithms,according to the characteristics of too much background noise in the images,the surface of the target object with regular shape has less noise,an Otsu threshold segmentation algorithm with median filtering combined with manual frame selection is proposed,this method can effectively eliminates the image noise caused by the high brightness of the projector when the camera is shooting and retains the edge information of the target object.3.In the investigation,we built an pipeline leakage detection system for industry,which uses the VGG16 model combined with the early stopping method,and compares experiments with a variety of deep learning-based algorithms to verify the effectiveness of the model,the training accuracy rate can reach 99.44%,the prediction accuracy rate reaches 97.0%,and the shortest single prediction time is about 0.2s,the best overall performance,this method's principle is relatively simple,easy to implement,and the cost of system construction is low,which can meet the detection requirements of industrial scenes.