Precision Inspection Theory And Technique Study For Micro Size Based On Computer Vision

Inspection technique based on computer vision is a modern inspection technique, which is rooting on modern optics and integrating optoelectronics, computer image, information processing, computer vision and other techniques. Over recent 30 years, computer vision detection has thrived with new concepts, theories and methods emerging continuously. It has been applied across industrial, agricultural, military, medical and other fields, and gained great progresses. However, from the perspective of development, computer vision detection technology is still in the progress itself, and many theories and algorithms remain to be improved. Vision inspection technique has not yet been applied broadly in industrial inspection, and contains potential that needs to be mined and utilized as soon as possible.At the present, industrial field has higher and higher requirements for micro size inspection precision and speed, which traditional inspection methods can hardly meet. In order to improve inspection precision and speed for micro size, theories and key technique of computer vision inspection are deeply analyzed and researched.The dissertation analyses theoretical and practical values of vision based inspection, as well as research status and trend of key computer vision inspection technique, and explains problems existing in and development trend of computer vision inspection. It also discusses the importance, categories, and performance assessment indicator of computer vision detection systems. Basic structure of the system is illustrated in detail. Configuration and selection rules of hardware in the system is discussed. It is explained that the software is important too. The dissertation also deals with typical methods for image noise elimination, including frequency-domain lowpass filtering, mean filtering, median filtering and edge-preserving filtering, as well as their respective advantages and disadvantages. It analyses threshold method and edge detection method for image segmentation, explains in detail histogram bimodal method, minimum-error segmentation method, OTSU, probability relaxation iteration and other typical threshold segmentation methods, as well as first order differential operator, second order differential operator, Canny operator and other typical edge detection methods. Advantages and disadvantages of various methods are also analyzed. Then, different operator edge detection methods are compared through experiments.As for image recognition theory, Hough transformation method for image area boundary recognition is analyzed. On the basis of traditional Hough transformation, random Hough transformation method and their respective advantages and disadvantages, the dissertation puts forward an improved random Hough transformation method for circle detection. The judgment using image edge gradient direction information and geometric primitive shape constraint helps reduce useless accumulation and calculation amount. Through experimental comparison, the improved random Hough transformation method for circle detection put forward by this dissertation, which can not only detect circles in the image accurately, but also increase calculation efficiency compared with traditional Hough transformation and random Hough transformation, even if the circle in the image is not full. A new pattern recognition method, Support Vector Machine method, is analyzed, and its basic principle and regression principle are explained. It is illustrated that Support Vector Machine features lots of unique advantages when solving problems of small-size samples, non linear and high-dimensional pattern recognition, which are also expanded to solve function fitting and other problems.Sub-pixel edge localization technology is used to acquire high-precision detection result. The dissertation also discusses typical sub-pixel localization methods, mainly including fitting method, moment method and interpolation method. Least Squares Support Vector Machine Regression principle is illustrated, and on the basis a new sub-pixel localization method is put forward, that is the sub-pixel localization method based on Least Squares Support Vector Machine Regression. According to characteristics of straight line and circle, the dissertation builds least squares support vector regression model for straight line and for circle respectively, compares regression accuracy through experiment, analyzes factors affecting regression accuracy, and verifies sub-pixel precision. Both theoretical analysis and experimental research indicate that the edge acquired by least squares support vector regression passes sub-pixel localization of edge pixel point, featuring relatively high localization precision and robustness against noise. Micro plastic gear vision detection system is developed. Taking micro size plastic double gear as the research object, images are processed through such steps as image gathering, image pre-processing, image segmentation, and profile extraction. Then, center hole diameter, roundness error and gear tooth defect are detected through random Hough transformation for detecting circle, sub-pixel circle center localization, defect detection, etc. Real-time image collecting is realized through triggering control on the camera by image gathering card settings. 2-step method is adopted for segmentation, and bi-threshold image segmentation method based on minimum-error segmentation is put forward, realizing rapidly and accurately image segmentation for big and small gears. The method is suitable for bi-threshold selection for similar types of parts. Improved random Hough transformation is used for integer pixel orientation of gear hole center, and the circle fitting based on Least Squares Support Vector Regression method is adopted for sub-pixel localization of circle center, and circle center is taken on as the orientation datum to make a virtual circle. The single pixel coordinates of insection points between virtual circle and gear contour are located by the polynominal interpolation method. Gear tooth defect detection is realized through virtual circle scanning. Experiment results show that detection precision is relatively high. The speed of detecting a single gear meets the requirement for online detection speed.Based on computer vision inspection technique, precision measurement is carried out for micro size. Taking the clip as the study object, the hardware structure of clip precision measurement system is built, and relevant measurement software developed. Micro sizes in a clip are detected. Firstly, the image is processed by image collecting, system calibration, image pre-processing, thresholding image segmentation, contour extraction and optimization. Secondly, sub-pixel localization based on Least Squares Support Vector Linear Regression is adopted to fit lines in the detected areas. According to the distance aquired, two line equations are determined. Thirdly, points of insection between two lines and line edge of the clip slot are determined separately, and the distance between points of insection are calculated. Experiment results show that the method proposed can enhance detection precision remarkably, and measurement limit error to average value is±0.3338μm.The dissertation also includes the development of vision detection software on Windows XP platform using Visual C++6.0. The measurement result indicates that the system software meets requirements of industrial detection.The last part concludes research findings and innovations in the dissertation, as well as suggestions on future work. The achievements of the dissertation are of theoretical significance and practical value for the development of image recognition theory, and will drive the application of computer vision in industrial quality detection.