Research On Weak Feature Material Defect Detection Technology Based On Machine Vision

With the development of industry,people have higher and higher requirements for the quality of the product surface.The material surface defects will affect the subsequent processing and sales of the product,is an important indicator to evaluate the product qualified or not.Material in the production process due to the impact of production materials,production equipment and other factors will appear scratches,cracks,chipping,scrapes,pits and other surface defects.At present,the common detection method for weak characteristics of material defects is manual detection.Because the defects are not obvious,usually need to be detected in the environment of playing strong light,quality inspectors in the long detection process,due to fatigue,occupational diseases and subjective factors,it is easy to cause mis-inspection and leakage.In response to these problems,this paper uses machine vision technology to study the weak feature material defect detection technology.The main research contents are as follows:(1)The weak feature defects are divided into two categories,one is the material defects with tiny features(minimum defect diameter of 0.03mm),which have the characteristics of small features,uniform background texture,and certain differences between defects and background;the other is the material defects with complex texture background.This paper analyzes the difficulty of detecting tiny features and complex textures.(2)A special detection algorithm is designed for the type of tiny defects in this paper.First,camera internal calibration is performed to reduce the error in sample acquisition and improve the accuracy of initial localization;then image pre-processing,edge detection and image segmentation are performed on the acquired sample images;after that,four vertices are positioned for each small glass piece to obtain the target rectangle.Since there will be missed detection in the localization detection,the blob algorithm is used to locate the vertices twice.After getting the coordinates of all small glass piece vertices,each small rectangle is keyed out by the mask image,and then the binarization process of local adaptive thresholding is performed for each small rectangle separately;finally,the detection and labeling of defects are performed by the binary image.(3)For multi-texture defect detection,a target detection algorithm based on improved YOLOv5 is used for defect identification.Firstly,the collected sample data sets are enhanced,mainly by single-sample data enhancement and multiple-sample data enhancement to expand the number of samples;then,based on the YOLOv5 model,we optimize and improve the YOLOv5 model by embedding attention mechanisms in the YOLOv5 model backbone network and neck network,respectively,to help the model better capture spatial and channel correlations when processing feature maps at different scales.In this way,the diversity and differentiation of features are enhanced,which helps the network to better distinguish the target and background and improve the detection of weak features.The impact of different location-embedded attention mechanisms on the detection accuracy of the dataset is analyzed to determine the optimal detection scheme.The detection accuracy can be improved by data augmentation and embedding the attention mechanism method in the network structure,which has better detection effect.(4)The upper computer development software based on QT is designed,and the glass sample defect detection task is deployed into the software to realize the detection of glass defects.