Image Classification Based On Laplacian Sparse Coding

Image classification is an important topic in computer vision and machine learning.Image classification is a method that classifies an image to a one of predefined categories by using certain properties.The process of image classification generally includes feature extraction and coding,feature fusion and classifier training.And the first two parts are combined into the image representation.Since there are many complicated and indescribable information in images,and there are wide differences between physical properties of images and information of human understanding,the issues of how to extract features from images effectively and appropriately become the key problems of image representation.The traditional sparse coding(SC)method is an effective method for image representation in recent years.Different sparse coding methods have solved the problems of lack of spatial information and visual vocabulary distribution in Bag of Words(BoW)model,and have improved image classification accuracy.However,these methods often encounter the two shortcomings: First,sparse coding is very sensitive to changes in features,which leads to the instability of the codes,thus similar features might be encoded into different codewords.Second,image representation and classification are relatively independent in these coding methods,so visual and semantic information of image might be lost.The combination of Laplacian regularization,locality and non-negativity can improve the instability of codes and ensure an independent encoding process.In addition,combining with the context information of image and comprehensively considering the image representation and classification make more semantic information and context information be preserved.In this paper,locality and non-negativity are introduced into the combined optimization of Laplacian sparse coding in order to improve the instability of codes to a great extent,which makes the similar features be encoded into similar codewords.Based on that,the context information is added to the reflection of the relationship between the semantic objects in the images and the surrounding environment of the properties.We jointly consider the image representation and classification to preserve more semantic information.Therefore,it can understand the images more adequately and effectively so as to achieve better image classification performance.The specific research work is described as follows:1.This dissertation has made improvements about the problem that the use of subtraction in the combined optimization problems of Laplacian sparse coding may make features offset each other.By adding a non-negative constraint for dictionary and codes,the combined optimization problems only involve operation of addition,which makes more local features and their relationship be preserved.2.The shortcoming that Laplacian sparse coding method ignores local information between features is improved in this.Inspired by locality-constrained linear coding,a local constraint for codes is added on the basis of non-negativity,thus one feature descriptor is represented by more basis from the codebook,which makes ensure that similar features share there local basis.Therefore,more local information between features is preserved.3.In order to obtain a pure dictionary that is more robust against noise,we introduce a proper sparseness for the dictionary.So the dictionary is equipped with some attributes such as orientation,bandpass,non-negativity,and localization,which results in part-based representation and improves the resulting decomposition.4.Image representation and classification are normally independent processes,which might cause the lack of context information of images.In this dissertation we improve this routine.After obtaining the Laplacian sparse coding by incorporating locality and non-negativity,we jointly consider the image representation and classification.We randomly select some codes of images to generate joint space based on context information and feature information and context information in the images are fused more effectively,which adequately considers the context relationship between image features in order to represent images more properly and effectively.