Research Of Image Classification Algorithm Based On Improved Locality-constrained Linear Coding

In the age of big data,we have tens of thousands of messages needed to be disposed of imminently every day.As an intuitive way to express information,images have been universally used in every walk of life.How to classify images quickly and efficiently is one of the hot topics that are worthy of exploration.The traditional manual classification requires a great deal of cost,and the classification effect may not be ideal.By combining computer technology with image classification technology,computers can recognize and classify images.That plays great significance.Among too much image classification technology,the bag-of-features model is a technical framework that is commonly used.It regards the image as a collection of feature vectors.To represent the image,it extracts the image features and then codes them.After that it chooses an appropriate classification algorithm to complete the classification.When we classify images,the features are important criteria.Feature coding has an important influence on the representation of images and the classification results.It is an important step in image classification.Locality-constrained Linear Coding(LLC)is a popular feature coding method.It takes full account of the locality of features.For each feature descriptor to be encoded,the LLC algorithm will select a number of similar visual words for it,to form a corresponding local coordinate system,and then reconstruct the feature descriptor in the local coordinate.The LLC algorithm has good reconstruction,local sparsity and smoothness;therefore it has been an important application in various fields.However,when selecting the K-nearest neighbors for the feature descriptor,the traditional LLC algorithm chooses the Euclidean distance as a unique metric.But it can't reflect the similarities between feature vectors very well.In addition,the information included indifferent regions of the image has different effects on the classification result.But the traditional LLC algorithm ignores the fact.When reconstructing features,it does not calculate the weight of each vector basis.Based on the above analysis,this paper has improved the traditional LLC,and proposes Locality-constrained Linear Coding Based on Hybrid Similarity and Weighted Bases(HSWLLC).The specific work is as follows:(1)The Euclidean distance and the histogram intersection are both common standard for the similarity measure of feature vectors.Nevertheless,the Euclidean distance can't compare vectors very well inherently.Therefore,using the Euclidean distance as the only measure has some limitations.Histogram intersection calculates the common part of two vectors on different components,which shows the similarities between vectors very well.Therefore,in this paper,we combine the two methods,assign different weights to them,and propose hybrid similarity.(2)In order to make full use of the important information in the image,and to reflect the effect of different regions on the classification result well,this paper gives different weights to vector bases.After local bases are selected for the feature to be encoded,we normalize the weights of local bases,to obtain the weight values of different bases in the local coordinate.In addition,during feature reconstruction,non-negative constrain is added to improve the stability of the algorithm in this paper.In order to verify the performance of our algorithm,experiments are carried out on three commonly used datasets.The method of this paper is compared with other methods,and the influence of some parameters on the classification result is also studied.The experimental results prove that the proposed algorithm not only improves the classification accuracy,but also enhances the stability of the coding.