Research On Defect Detection Method Of Glass Cover Based On Machine Vision

In today's intelligent equipment into thousands of households,the demand for glass cover as an important part of the display equipment is increasing day by day.The huge market stock and the potential of future growth are the focus for many countries and enterprises to compete for.As the last process of glass factory,inspection not only needs to ensure the quality of the product,but also needs to ensure that the time spent meets the requirements of the production line.Combined with the existing technology,this paper improves the original industrial image recognition algorithm,and designs a glass defect detection system based on multi-channel acquisition and comprehensive evaluation.The main research of this paper is as follows:(1)In the glass cover defects detection,this paper industrial image and the difference of natural images,analyzes its performance,color gray level distribution and image size and other characteristics,in order to meet the requirements of production line and high precision,low latency,this scheme has adopted a pre-segmentation and defect discrimination of flaw detection scheme,the scheme of the pre-segmentation steps using template matching algorithm based on difference technology,expounds the existing registration algorithms in industry stability and robustness of target detection,and on the basis of the existing algorithm proposed a registration algorithm based on geometric structure,By collecting the centroid,target contour and centroid of the template image and sample image respectively,the minimum structure tree is constructed respectively,and the mapping model from the sample image to the template image is constructed through the iterative optimization method,so as to complete the subtracting process.(2)In this paper,a multi-scale integrated residuals convolutional neural network is proposed on the basis of comprehensive industry solutions.Compared with the traditional classification model,the proposed scheme not only requires the accurate expression of small features,but also requires the model to reasonably represent the overall information of the image.Although the characterization ability is improved when the number of convolution layers deepens,the algorithm complexity will also increase exponentially,and there is a risk of over-fitting.Therefore,the model proposed in this paper increases the width of the convolutional layer while limiting the depth of the neural network.This structure can not only accurately describe the contour information of defects,but also improve its characterization ability for small features,so as to accurately judge the features of small defects in the sample.(3)This paper modeled different processes,analyzed the advantages and disadvantages of image processing schemes based on structural optics and traditional industrial image processing schemes,and then proposed a multi-channel defect information fusion algorithm.Firstly,the mapping model between different optical devices is obtained in the decoded output part of the acquisition end,and then different prediction results are obtained after various segmentation algorithms.The prediction results are mapped to the same coordinates by the mapping model for distance judgment,and then the fuzzy k-means algorithm is used for clustering,so as to get the final result.(4)In this paper,a software system with "high cohesion and low coupling" is designed based on the above technical points.Firstly,by building middleware,the strong coupling relationship between service components and hardware devices is weakened,so as to realize a system that can collect the on-site decision results,and then store them in the background data set after manual inspection.Based on this data set,not only can a more efficient algorithm model be trained,but also the accuracy and availability of the current deployed algorithm can be evaluated by tracking the historical error record.