Study Of Machine Vision Detection Based On Tiny Components In Middle Scale

With the increasing of miniature components and devices in recent years,the demand for the detection of these parts are increasing synchronously at the same time.To meet the demand,with the help of Zhejiang Provincial Natural Science Foundation Project(LY13E050026)-Research on the key technology in inspection of three-dimensional micro components of medium scale components,this paper do study on the meso-scale components inspection theories,including image denoising,edge detection and sub-pixel edge location and so on,and for the inspection in SMD chip,hardware platform was set up in machine visual detection and the visual inspection software system was developed,eliminating the unqualified chip and determining the polarity,center position and rotation angle of eligibility chip.Imaging system exists phenomena such as uneven illumination and lens distortion problems.For meso-scale components image detection system,because of the factors of its small view field and multi-shot,compared to the macro vision,the effect brought to the images of thees two problems is more apparent.In this paper,the uneven illumination correction and lens distortion correction were carried on for the existing experimental detection system.Adopt a uniform gray scale value calibration to calibrate the uneven illumination and the calibration nonuniform coefficient is determined by the plate grayscale and gray average.Select a circular grid calibration plate to calibrate the lens distortion and get the distortion coefficient by determining the ideal location and the actual position of the grid.The actual position is the deviation from the ideal position deviation due to the lens distortion.Much noise generates during the actual testing process of small components.In this paper,a chip image was got in the lab environment and simulate the actual chip image by adding noise.And then make research on the meso-scale image with salt and pepper noise.In less valid signal condition,make full use of existing information retrieval noise images which includes distribution information-gray histogram,neighborhood information-spatial proximity and grayscale correlation.With them as weights,firstly use adaptive weighted mean filte,then adaptive mean filtering to reduce noise.Experiments show that the algorithm works well in denosing salt and pepper noise.For the situations of meso-scale micro components edge profile complex,much noise and detail fuzzy,an improved Canny detection algorithm was proposed.The algorithm focuses on the improvement of threshold denoising and dual-threshold.Denoising includes adaptive median filter to remove salt and pepper noise and adaptive gaussian filter to remove gaussian noise.This paperfocuses on the selection of the scale factor.In Dual-threshold section,for weak edge detection,by cutting image into blocks to select threshold,and do edge detection.The threshold selection including roughing and handpick selection.Sorting the image block pixel gradient values,find the roughing threshold.Then the exact value was determined by the gradient of the histogram graph.In meso-scale components detection area,the contrast experiments prove this method do better than traditional Canny detection.To make further improvement in system accuracy,the paper also do further research on sub-pixel positioning for the edges(lines,circles,arcs)which alwalys existing in middle-scale image.when the edge pixel level position was known,the edge normal direction was obtained through the Hough transform,and the grayscale difference which includes forward difference and backward difference were got through calculating the gray value of interpolation points in the normal direction.Center position can be figured out by treating the normalized difference value as the position weight,then the ideal edge positions was got.The ideal edge could be got based on the edge curve shape of the least-squares fitting.The experiments prove that the algorithm of subpixel localization accuracy is high.Finally,software system for SMT chip detection was developed.Through statistical sample characteristics,the parameter index values were received including area,rectangular degree and the ratio of long and short sides to estimate positive or negative,qualification or un-qualification.these date were treated as the testing standard when we do detection in SMD.The unqualified chips can be figured out,and to the qualified ones,the polarity and center position and the rotation angle information were output,which help robot to do next steps.The measurement time is less than 1s,achieving real time detection requirements.