The Research On Gear Appearance Defect Detection Method Based On Machine Vision

Gears are widely used in mechanical products,and their quality directly affects the life,transmission ratio and motion accuracy of mechanical moving parts.At present,the detection of gear appearance defects mainly depends on manual sampling,which can not meet the requirements of mass production of gears.At the same time,manual inspection method has the disadvantage of poor reliability,so it is urgent to carry out automatic gear appearance defect detection method research.In this paper,the key problems of gear appearance detection,such as inaccurate segmentation of gear image,gear tooth positioning and low detection accuracy,are deeply studied,and a gear appearance defect detection technology based on machine vision is determined.The main research contents and results are as follows:(1)To solve the problems of gear image segmentation and ROI positioning,a GA-Ostu image threshold segmentation method for gear and an improved A-KAZE and feature contribution ROI positioning method for gear image are proposed.An Ostu algorithm which combines genetic algorithm search strategy is studied.This method divides background from target,then uses mask operation to locate ROI preliminarily,and finally uses improved A-KAZE for feature detection and matching in the area of preliminary positioning,and obtains homography matrix by using characteristic contribution degree and error matching point of RANSAC filtering,and calculates reference image in original position.The position and size of the map enable accurate positioning of the ROI area(gear tooth area),the correct matching rate of feature points is 97.99%,while the ROI positioning speed is only 1.137 s.(2)Analyzing the problems existing in gear appearance defect detection and studying a gear appearance defect detection model based on improved Faster R-CNN.Referring to the structure idea of "jump connection" in residual network,the basic convolution neural network VGNet-16 is improved on the existing network structure to enhance the extraction ability of network features.mPA is 1.51 percentage points higher than VGNet-16.A new loss function AMF-Softmax is formed by combining the unbalanced evaluation index F-score with loss function AM-Softmax and setting corresponding conditions,and the mPA reaches 96.14%.(3)Building the gear appearance detection experimental platform and software test environment.In this experimental platform,the above theoretical methods are verified and analyzed by using the gear appearance data collected in the field.In the experiments of image segmentation and ROI positioning,the results of image threshold segmentation,several feature detections and matching and ROI positioning comparison are given.In the gear defect detection experiment,the comparison results of different target detection network models Faster R-CNN based on convolution neural network,different loss functions and improved loss function AMF-Softmax are given.