Method And Application Of Image Feature Extraction Research Based On Representation And Projection Learning

Feature extraction can simplify high-dimensional data,retain valuable information,and reduce the subsequent image processing workload.The research work in this paper focuses on image feature extraction algorithms based on representation and projection learning.At present,many existing feature extraction algorithms have problems such as stale algorithm theory,high complexity and outstanding performance.We need to come up with new approaches,then the feature extraction ability of the model will be improved.Thanks to our efforts,new feature extraction model and optimization algorithm were proposed and it has been verified with a large number of experiments.The specific research work is as follows:First,we propose an image feature extraction algorithm based on orthogonal projection learning.Relevant scholars have proposed many excellent algorithms in terms of projection reduction,but there are still some unsolved problems.Such us focusing only on the data in the embedded space ignores the projection process from the original space to the embedded space and unable to make full use of the label of the sample to know the projection process,etc.Based on the above problems,an image feature extraction algorithm based on orthogonal projection learning was been proposed.It has three main advantages:(1)Using norm constraints,the main information of the sample is embedded in the low-dimensional space;(2)The classification loss function is introduced to effectively use the label information of the sample;(3)The weighted Schatten-p norm is used to constrain the rank.Experimental results show that the proposed mathematical model can extract the main feature information of the image.Second,we propose an image feature extraction algorithm based on joint kernel projection and representation learning.Image information tends to be complex and nonlinear in the raw data space.The existing image feature extraction algorithm usually builds a mathematical model based on the original data space.Then Projection is performed using the learned projection matrix and completes the feature extraction task.The mapping algorithm based on kernel trick maps the image data in the original space to the Hilbert space through the kernel function,which is more conducive to finding the feature relationship of the data in the high-dimensional space.Real-life data is often non-linear,which makes it difficult to classify and identify data.For this reason,through the introduction of kernel techniques,the nonlinear data in the original space is projected into the Hilbert space.Then the feature relationship is learned in high-dimensional space,which can guide the low-rank representation learning process of image data.This manner will get simpler and more efficient low-dimensional feature information.Finally,we design a face recognition software based on the two algorithms mentioned above,and successfully apply the proposed feature extraction algorithm to the face classification recognition task.Based on the existing algorithms,this paper analyzes its performance,and then proposes new model,and conducts relevant experimental verification on multiple data sets.The results show that the algorithm proposed in this paper is advanced.In addition,we applied the proposed algorithm to face image classification,and got good classification results,which proved the practicability of the algorithm.