Research On Machine Vision Image Non-Negative Feature Extraction Approach In Multiple-Constraint Conditions

In recent years,machine vision has become an important branch of the rapid development of artificial intelligence.Generally,machine vision is to use machines to replace human eyes for measurement and judgment.The image sensor device takes the target and covert into the image,sends it to the dedicated image processing system,and obtains the subject target shape information.Then,a variety of operations are done for extracting the target feature and judging the status,category and other information.Therefore,image feature extraction becomes one of the core technologies of machine vision image processing related research works.With the continuous improvement of machine vision theory,more and more image feature extraction methods start from different points of view and get more distinctive image features while retaining the original information of the image possibly.Dimensionality reduction becomes one of the most popular directions for image feature extraction because of the completeness of its mathematical theory and the convenience of realization.The dimensionality reduction based image feature extraction can map the high dimensional image to the lower dimensional space and get the image distinguishing feature.In particular,the features obtained by non-negative constrained dimensionality reduction can describe the purely additivite and sparse image features,and can explain the global and local relationship,which is in line with the intuitive image understanding.Therefore,the image non-negative feature extraction method based on Non-negative Matrix Factorization(NMF)plays an important role in various applications of image processing.However,there are still some problems in the existing methods of image non-negative feature extraction in multiple-constraint conditions.Therefore,this paper focuses on the characteristics of image data and studies non-negative feature extraction of machine vision in multiple-constraint conditions.Image non-negative feature extraction approaches need to consider a variety of image data,mainly including images without labels,images with labels but non-linear correlation,and few images with labels.For this.this paper introduces several constraints to solve the problems of image non-negative feature extraction under these circumstances,and conducts method verification on the large-scale image data sets.The main research contents and achievements of this paper can be summarized as follows:(1)As for the images without labels,an image non-negative feature extraction method based on topological non-negative matrix factorization is proposed.The existing image non-negative feature extraction methods based on NMF set the rank as the number of classes,and then uses clustering algorithm to verify the performance of feature description.This destroys the local geometry among the images.The topological non-negative matrix factorization model introduces two constraints of rank relaxation and image feature graph structure,and integrates two tasks of image feature extraction and clustering into one optimization objective function to obtain non-negative features that can express the topological relation of images.By comparing with other non-negative matrix factorization methods on the data sets of ORL,UMIST,MNIST,COIL20 and COIL 100,this method has obtained the best ACC and NMI.Compared with all the compared methods,the proposed method improves 1%-4%and 1%-5%in terms of ACC and NMI on the experimental datasets,respectively.(2)An image non-negative feature extraction method based on label embedded kernel non-negative matrix factorization for nonlinear-related labeled images is proposed.For labeled images,the existing NMF methods adopt Fisher or maximum margin criterion to extract the feature using image sample mean,which is highly dependent on image sample distribution.In the proposed objective function,the non-negative matrix factorization model introduces two constraints,label embedded one and kernel function mapping,which does not assume the image sample distribution,and can directly obtain the discriminant image non-negative features by using the nonlinear relationship among the images.By comparing with other common methods on the data sets of ORL,YALE,MNIST and COIL20,this method achieves the best recognition rate.Besides,this method improves 2%+compared with the second best compared method.Typically,the proposed method performs better on the MINST dataset which consists of more complex images and improves 5%compared with the best result of the other methods.(3)A linear regression combined with semi-supervised non-negative matrix factorization is proposed for extracting non-negative features from a small number of labeled and a large number of unlabeled images.The common semi-supervised NMF method introduces the graph structure of original images in the process of matrix factorization for feature extraction,which is sensitive to noise interference.In addition,they can only obtain the non-negative features of the training images,and cannot obtain the non-negative features of the new images.The proposed method introduces three constraints,which are label propagation,self-organized graph structure and linear regression constraints on non-negative image features.Then,this method can obtain the non-negative features of training images,category attributes,projection matrix for extracting new image features,and the parameters of category prediction.By comparing the experimental results of the proposed approach with those of common clustering and semi-supervised classification methods on the data sets of MNIST,USPS,ORL,YALE,UMIST,COIL20,COIL 100 and DTD,this method can obtain the best ACC and NMI.As for the recognition of unlabeled images and testing images,the proposed method improves 2%-11%and 1%-3%in terms of correct recognition rate compared to the best employed methods,respectively.The works above mainly focus on the non-negative feature extraction of static images in machine vision.Therefore,this paper further studies the non-negative feature extraction method of dynamic images in machine vision,and proposes a dynamic image non-negative feature extraction method based on local subspace learning.For dynamic images,common methods assume model for each image set,and parameter estimation is not accurate.By learning a non-negative sparse map,this method introduces the constraint conditions of inter-and intra-class in the local set to obtain the non-negative features of the dynamic image sequence and reduce the computational complexity.By comparing the experimental results of approaches on the image set data sets of Honda,YouTube Celebrities,CMU Mobo and ETH-80,this method achieves the highest correct recognition rate and improves 1%compared with the best one of the employed methods.The time cost of the proposed method is less than most of the compared methods.