Research On Image Classification Based On Multi-feature Fusion

With the development of information technology and Internet technology,the image information generated by every day has increased explosively.In the face of so many images,we want to know what information the image expresses,which produces the image classification technology.Image classification technology is now applied to all aspects of life,and it has great research significance and value,a good image classification algorithm can reduce the manual classification to a large extent burden,fast image processing,has now become one of the important research directions in the field of computer vision.In this paper,based on the traditional image classification framework,a multi feature fusion image classification algorithm based on significant detection and target location is proposed.The traditional image classification method usually extracts a single feature to classify the image,and all the images are directly operated on the images in the data set.The general process of the traditional image classification algorithm:firstly,extract the local low-level features;then,using the training set image local training low-level feature size of K visual dictionary;then,the extracted low-level features using feature encoding method and visual dictionary encoding;finally,training a classifier for image classification.This article is mainly done in the following aspects:(1)Image preprocessing,based on the SEG significance detection algorithm,the segmentation and location algorithm of image target area is proposed.First of all,the use of SEG saliency detection algorithm was the image of the original image,namely the probability map;then,using Otsu threshold segmentation algorithm for image segmentation was obtained,two value image,noise points and knowledge of the use of morphological processing to remove fine;after finding the candidate target region,the proposed method this paper is to find the largest connected region is greater than or equal to the 1/10 area as the target area;finally,according to the selected candidate target region to complete target area.(2)Three different types of underlying features are extracted.Because a single feature is too weak for describing the image,it is not enough to describe all the feature information of the image,so this paper applies many different types of features to describe the image information.This paper mainly extracts three underlying features:dense SIFT features,HOG features and Garbor features.The difference between the dense SIFT feature and the traditional SIFT feature is that the image is densely grid sampled,and then the SIFT feature in the grid is extracted,so that both local information and global information can be extracted.The Garbor feature is achieved by filtering the filtered image obtained from the evenly divided image blocks and the filters with different scales and different directions.Compared with the directly filtered images,the dimension is less and the computation speed is increased.(3)The encoding of the underlying features and the extraction of the spatial Pyramid pooling features are realized.The underlying feature is coded because of its weak distinguishing ability and high feature dimension.First of all,the use of online clustering K-Means clustering algorithm to extract the low-level features,the formation of different size dictionary codebook;then,based on analyzing the existing encoding method,select the local linear encoding reconstruction error is relatively small for encoding the underlying features;after using the maximum pool,get the maximum pool characteristics image;finally,using spatial Pyramid model,the image is divided in different resolution,the extraction of each image block pool features,get features with spatial information encoding.(4)A fusion method based on multiple features is proposed.This paper uses the fusion method and fusion method for later experiments,in addition to a number of features of direct series mosaic,also put forward according to the recognition rate of single feature method to determine the feature weight fusion;when the late fusion,this idea is:three different encoding feature set from training then,input to the three SVM classifiers,respectively three training SVM classifier for test images,respectively.The three features are input into the trained classifier,get the results according to the comprehensive score and label the output decision.Through the experiment comparison,the fusion method with direct series splicing is better,so the contrast experiment uses this fusion method to do the contrast experiment.In this paper,we do two sets of comparative experiments on the open data set Caltech101,and compare it with the traditional image classification method and single feature classification method,and verify the feasibility and accuracy of the algorithm.