Research On Dual Space-constrained Non-negative Matrix Factorization Method For Surface Defect Detection Of Mechanical Parts

With the rapid development of production technology,mechanical parts have not only increased in production,but also continuously improved in quality.However,quality problems still exist,such as surface defects,shape and size problems.Nowadays,machine vision inspection technology is developing rapidly and gradually used in the detection of parts surface defects.Meanwhile,the inspection speed has been greatly improved.In the traditional surface defect detection process,the dimensionality of the collected target image is high,which cannot effectively capture the characteristic information of the image.This paper mainly studies how to discover the low-dimensional features of the original high-dimensional mechanical parts image data to improve the performance of cluster and classification.The traditional non-negative matrix factorization methods cannot not fully consider the prior information of the samples,such as the structure information of the data space,the manifold structure of the feature space,and the category information of the sample.In addition,due to the correlation of features,traditional NMF using Euclidean distance cannot effectively measure the reconstruction error between samples.To deal with these issues,this article conducts in-depth research:1.Two self-made surface defect detection data sets of mechanical parts,including gear surface data set and sliding rail surface data set.In addition,the steel plate image library on the Internet is processed to form a data set suitable for the method in this paper.2.A dual local learning regularized non-negative matrix factorization(DLLNMF)algorithm is proposed.First of all,this method not only considers the local structure information of the data space by constructing a local learning regularizer in data space,but also considers the local structure information of the feature space by constructing a local learning regularizer between the features.Then the dual local learning regularization term is integrated into the traditional NMF model to fully explore the geometric manifold structure information and discrimination information of the image data.In addition,we use the known category information in the sample as a hard constraint to further propose the Dual Local Learning regularized Non-negative Matrix Factorization(DLLCNMF)algorithm on the basis of the DLLNMF algorithm.Experiments on three surface defect datasets show that the DLLNMF and DLLCNMF algorithms can effectively improve the clustering performance.3.A dual graph regularized Non-negative Matrix Factorization with Sinkhorn Distance(DSDNMF)algorithm based on Sinkhorn distance is proposed.Because the traditional NMF method using Euclidean distance cannot effectively measure the reconstruction error between samples,this method uses Sinkhorn distance instead of the traditional Euclidean distance.In addition,we adopt the dual Laplacian graphs to construct a regularization item to preserve the manifold structure information of the data in this method.It also preserves the manifold structure information of the sample and explores the intrinsic geometric structure relationships between the samples.Experiments on three surface defect data sets show that DSDNMF can effectively measure the reconstruction error between samples and achieve better clustering performance.