Research On Machine Vision-based Detection Technology For Surface Defects Of Micro-hole Structures

With the progress of the manufacturing industry,there are more and more equipment with special structures.Among them,the micro-hole structure is extensively applied to the fields of automobiles,energy,chemical industry,national defense,etc.However,insufficient manufacturing process and device fatigue may cause defects such as cracks and pores in the inner surface of these holes.It will heavily hinder the performance of the equipment and even lead to safety accidents.Thus,the quality of those structure is vitally important.Limited by the size of the small hole,it is difficult for the inspection equipment to obtain the image of the inner surface.Besides,the process of defect detection should be intelligent to meet the requirement of online measurement.Therefore,it is valuable to study the inner surface detection system of micro-hole structure and its related algorithms.The main work of this paper is as follows:1.A new detection principle of importing external illumination and exporting the image of the inner surface based on sight-pipe is proposed.The optical measurement system,motion control system and its supporting software are designed based on the principal.An engineering prototype is built to detect micro-hole with an inner diameter of 4-8mm and a depth of no more than 45 mm.2.The relationship between the image of the inner surface and the position of the system is analyzed.An algorithm for extracting effective region is proposed to shield the obstacle and establishing a mapping between the position system and the measured part.The obtained effective region can be applied to feature pattern measurement,panoramic image stitching and defect detection.A calibration algorithm is designed based on the optical resolution calibrator.With the cooperation of the image enhancement algorithm,the calibrated system achieved quantitative measurement.3.Based on the deep convolutional network,a discriminant network and a segmentation network are proposed.The former is responsible for judging whether there are defects or not,the latter is introduced to predict the contour of the defects.The accuracy and mean intersection of union of the overall network on the dataset achieved98.5% and 0.834 respectively,which realized pixel-level defect prediction in real time.4.The prototype of the detection system and its corresponding algorithms have been integrated as micro-hole structure detector.The work principles and the proposed algorithms in this paper have been verified in experiment.