Research On Image Classification Based On Gaussian Distribution Modeling

Image classification is one of the most important and fundamental topics in computer vision community,which has attracted a lot of attentions in recent decades.With the rapid development of image classification technology as well as the strong demand for artificial intelligence,products based on image classification teem in intelligent life,security and protection and medical diagnostics,and so on.However,robust and accurate image classification is still a very challenging task because objects are diverse and variety in real world,and a variety of noise factors have bad influences on imaging.Probability distribution modeling is strong representation method,and is robust to partial information loss and noise.However,such methods have special structures and suffer from high complexity,resulting in unsatisfactory performance in image classification tasks.The goal of this dissertation is to develop robust and accurate image classification methods based on Gaussian distribution modeling by sovling above problems and taking full advantage of its modeling ability.To this end,contributions of dissertation are concluded as the following four parts.(1)The key of image classification based on Gaussian distribution modeling is fully exploiting their special structures.Because the space of Gaussian distribution forms an unknown Riemannian manifold,the traditional operations in linear space cannot be directly applied to the Gaussian manifold.For reasonably and effectively exploiting Gaussian distribution modeling,this dissertation studies and analyzes the space of Gaussian distribution.For this purpose,through rigorous mathematical proof,this dissertation reveals the space of Gaussian distribution equips with a Lie group structure,and proposes two kinds of novel Gaussian embedding methods based on the theory of Lie group.These two kinds of methods map Gaussian distribution into the linear space for efficient usage while considering their both algebraic and geometric structures.The experimental results show that the proposed Gaussian embedding methods are superior to the existing methods,which builds the theoretical foundation for the usage of Gaussian distribution modeling.(2)This dissertation proposes a novel codebookless model based on Gaussian distribution modeling based on the above analysis and understanding of Gaussian manifold.Different from previous classification methods based on codebookless model,the proposed method utilizes Gaussian distribution modeling with full consideration of the special structure of Gaussian distribution,and Gaussian distribution is embedded into a linear space for efficient and effective classification.This method avoids the limitations caused by codebook in the popular Bag of Visual Words models,and it is among the first who shows that such codebookless model based on Gaussian distribution modeling is very competitive alternative to the mainstream Bag of Visual Words models.Meanwhile,this dissertation makes an experimental analysis for the proposed method,and concludes that more powerful local descriptors can bring more improvement for codebookless model than Bag of Visual Words models.(3)Because codebookless model based on Gaussian distribution modeling is more sensitive to local features,so this dissertation proposes a robust approximate infinite dimensional Gaussian descriptor based on above codebookless model.This Gaussian descriptor first employs features from deep convolutional neural networks and explicit feature mappings for approximating infinite dimension to enhance local features.Considering features from deep convolutional neural networks usual are of high dimsension and small sample size,so this dissertation proposes a novel regularized maximum likelihood estimation method to robustly estimate Gaussian in case of high dimension and small sample.This method can significantly improve performance of image classification method based on Gaussian distribution modeling.Moreover,it shows combination of Gaussian distribution modeling with deep convolutional neural networks is possible,and demonstrates the importance and necessity of robust estimation for Gaussian modeling in case of high dimension and small sample.(4)The robust approximate infinite dimensional Gaussian descriptor shows combination of Gaussian distribution modeling with deep convolutional neural networks can significantly enhance the modeling ability of Gaussian distribution.However,the aforementioned image classification methods based on Gaussian distribution modeling handle feature extraction,gaussian modeling and learning classifier in a separated manner,this dissertation proposes a novel global Gaussian distribution embedding network,which plugs a Gaussian distribution as image representation into deep convolutional neural network in an end-to-end learning manner.It achieves the goal of joint optimization of feature extraction,Gaussian modeling and classifier.This method designs a unified framework for image classification methods based on Gaussian distribution modeling,which further enhances the performance of the image classification methods based on Gaussian distribution modeling.To evaluate the effectiveness of the proposed methods,the dissertation applies these methods to various image classification tasks,including object recognition,scene categorization,texture/material image classification,large-scale image region classification,fine grained image recognition.The experimental results on 16 image benchmarks(containing more than millions of natural images from thousands of categories)show that image classification methods based on Gaussian distribution modeling are very accurate and robust,and outperforms their counterparts.