SIFT Feature Compression Algorithm Based On Compressive Sensing And Its Application In Image Classification

Image classification is the basic problem of computer vision.With the populari-zation of mobile Internet era,more and more mobile devices and wearable devices are required to be used for image classification.How to classify the image on the equip-ment with limited computing resources and limited space resources is very urgent.This paper has done some research work on it.Image feature extraction produces a large number of high-dimensional feature data,which can cause pressure on system space storage.The conventional solution is to leave the image to the server for feature extraction.However,for some special ap-plication scenarios such as mobile phone album classification,we can not upload user photos to the server in violation of user privacy.We can only rely on the client for feature extraction.In addition,the traditional feature reduction method reduces the characteristic data,but also changes the original property of the image feature,result-ing in the loss of characteristic information.In order to solve such problems better,this paper proposes an image classification method based on compressed sensing fea-ture compression coding,which includes:Firstly,the compression coding algorithm based on compressive sensing is pro-posed.The algoritlm firstly to SIFT characteristics of compressive sensing signal re-construction transform it into high-dimensional sparse signal,and then use sparse high-dimensional signals compressed coding,eventually will SIFT feature compres-sion for low dimensional data representation.The algorithm can recover the original characteristic information by decompressing.The algorithm can retain more feature information than feature reduction and recovery.Secondly,in order to reduce the time of feature compression,this paper proposes a new compressive sensing reconstruction algorithm:the threshold adaptive seg-mented orthogonal matching tracking(TaStOMP)algorithm.In each iteration,the TaStOMP algorithm sets threshold threshold according to residuals,and updates mul-tiple supporting atoms to accelerate the convergence rate during an iterative process.The experimental results show that the SIFT feature compression coding algorithm based on the TaStOMP algorithm has higher accuracy and less feature compression time than other compressive sensing reconstruction algorithms.Thirdly,combined with the feature of compression perception compression cod-ing algorithm to optimize the traditional image classification process and stage in the image feature extraction of image features extraction and compression,avoid seriali-zation feature file compression process,features can greatly reduce the storage space.The experimental data show that this method can reduce the image feature data to 40%and still keep the original image classification accuracy.