Application Of Three Improved Linear Regression Classification Methods In Face Image Recognition And Remote Sensing Image Scene Classification

Due to the development of artificial intelligence,computer technology,machine vision and many other factors,image processing technology has become one of the hot research spots and has received extensive attention from scholars at home and abroad.Image processing mainly includes image detection,image pre-processing,image extraction and image recognition and classification,among which image recognition and classification is the key step in image processing.Deep learning excels in image recognition and classification,but it requires a huge number of samples to train the model.However,sufficient samples are not available under some specific conditions,and in addition,the cost of obtaining samples and assigning labels is huge.For example,massive images of terrorists are not available,and the high cost and difficulty of remote sensing images make it difficult to acquire a large number of samples.Linear regression classification is an efficient classification method,which belongs to nearest neighbor subspace classification,and is widely used in face image recognition and remote sensing image scene classification.This method usually deals with the original high-dimensional images directly,but the original data contains a large amount of noise,redundant information,several separate clusters will exist in the same class.The three situations will reduce the effectiveness of linear regression classification.In order to solve the three problems,this thesis uses an improved linear regression classification methods to identify and classify images based on principal component analysis and IGG weight functions,local Fisher discriminant analysis,and local ideas.The main research is as follows:(1)A robust linear regression classification method based on principal component analysis and IGG weight function is proposed to solve the problems of poor under-sampling dimensionality reduction and weak robustness of linear regression classification.Principal component analysis is used to reduce the dimension of the high-dimensional image vector,and then the least squares method based on IGG weight function is used to estimate the linear regression coefficients in linear regression classification.The experimental results show that the method has better recognition and classification results compared with linear regression classification and robust linear regression classification based on IGG weight function regardless of whether noise is added or not.(2)A regularized least squares collaborative representation classification method based on local Fisher discriminant analysis is proposed to solve the problems that the dimensionality reduction effect of principal component analysis needs to be improved and the pseudo-inverse of the coefficients in solving collaborative representation classification.The local Fisher discriminant analysis is used to reduce the dimensionality of highdimensional images,and the parameters in the collaborative representation classification are estimated by the regularized least squares method.The experimental results show that this method has the highest recognition rate and classification accuracy compared with the four methods,i.e linear regression classification,linear regression classification based on principal component analysis,local Fisher discriminant analysis and regularized least squares collaborative representation classification.(3)The recognition rate of linear discriminant regression classification will be affected when several separate clusters exist in a certain class.In order to solve this problem,a linear local discriminant regression classification method is proposed.The high-dimensional image vectors are dimensionally reduced using principal component analysis,and then the reduceddimensional images are classified using linear local discriminant regression classification.The experimental results show that the recognition and classification results of this method are optimal compared with the other six methods,regardless of face images,object images or remote sensing images.