Research On Visual Inspection Of Surface Defects Of Milling Workpiece Based On Super Pixel

Visual defect detection is widely used in many industrial fields.Specifically,the application detecting surface defects for workpieces tackles the problem in actual production.On the one hand,the automatic precision control of the equipment is achieved to improve the efficiency of production.On the other hand,the problems are solved about offline operation and dependence on labor.In this thesis,we proposed a defect detection method using machine vision based on superpixel.Compared to traditional detections,this method has proven to be more effective against extracting defect informations and avoiding the influence of surface cutting paths during metal milling.This thesis consists of four parts.In the first part,we aim to cut workpieces and obtain informations about image and roughness of surface.According to the factors determining surface quality,the orthogonal experiment is constructed and completed by the multi-machining parameters,the various tools and the various materials.It obtains the rich information about surface topography.We also collect the images and measure roughness for these surfaces of the machined workpieces.In the second part of this thesis,we investigate how we can identify suspected defect areas based on the texture features of surface roughness.For the high-resolution,large-frame images captured on the surface of the workpiece,the image size is cut to reduce the data dimension.The sub-images are transformed into corresponding gray level co-occurrence matrices,and their characteristic parameters are extracted.Combining the measured roughness values,the mapping relationship is constructed.Comparing the parameters of the sub-images that include defects in the sample set,we define the discriminant criteria which screen sub-images containing suspicious defects.A defect image library is saved and created for further processing.In the third part,we focus on surface defects segmentation.How to determine each input parameter in SLIC algorithm is the key problem during the actual segmentation.We propose an adaptive superpixel segmentation algorithm combining with the extreme learning machine.The superpixel distance function is constructed by extracting the grayscale distance and spatial distance between the superpixels.We find the boundary between the defect region and the background region to aggregate the super-pixels of the defect and background area based on the DBSCAN clustering method.The logical binary images are obtained.In the forth part of this thesis,we build SVM classification model extracting geometric feature parameter from the binary image.The principal component analysis method is used to reduce the feature space of the feature vector.According to the characteristics of the defects generated by the milling method,the three classifier of PBT SVMs is designed for classification of defects.We implemented and demonstrated these ideas in the detection experiment of defects on the surface of milled workpieces.The results show the effectiveness of these approaches.Considering the edge information and the pixel space relationship of the defect area,the relatively complete defect area and its characteristic parameters are effectively extracted.At the same time,the results of defect classification is completed efficiently and accurately.We believe that this method provides a new angle and treatment for the complete segmentation defects from background texture in the surface defect detection of workpieces.