The Micro Defect Detection Of Stencil Based On Machine Vision

With the rapid development of industrial automation,the dependence of the production and processing industry on machine vision systems is increasing.Machine vision system is widely used in industrial field because of its high resolution and low distortion in image acquisition,high precision and low time consumption in image processing.The quality of stencil,the key abrasive tool used in the first process of SMT production line,has a direct influence on the quality of the subsequent processed products,so it is very necessary to detect its defects.In the existing defect detection schemes,except manual vision inspection,nearly all rely on the assistance of Gerber data files,and there are some shortcomings such as time-consuming,complex detection process,and poor detection effect.In this topic,by searching for common corner features of stencil defects,without using the Gerber data files,the extracted corner areas are classified to detect the defects of various holes on the stencil and locate them.The feedback data is used for the reprocessing of stencils to ensure the orderly execution of subsequent SMT production processes.After image preprocessing such as image stitching,hole partitioning,hole internal and external segmentation is performed on the captured stencil images,the Harris algorithm is used to extract the corner points.Then,features of the extracted corner regions are analyzed and the feature vectors of the region images are constructed.After that,the feature vectors are sent to the MLP neural network for training.And the MLP classification model is continuously optimized using feature normalization,PCA dimensionality reduction,and addition of regular terms.Finally,the average accuracy of the MLP model in the corner region classification on the test set is more than 95%,and the minimum accuracy is close to 92%.The model can quickly detect micro defects such as burrs,covering,blocking on the stencil holes,and the detection precision can be achieved 30?m on the calibration target,and it takes a short time.It solves the problems of low precision,poor real-time performance and high misjudgment rate in defect detection.