Visual Perception Modeling Based Local Image Structure Representation And Its Applications

After billions of years of evolution,mammalian has highly-developed vision systems.So researching the vision mechanisms of mammalian is meaningful for both of science and engineering.The goal of computer vision is to develop vision systems that have similar functions with the vision systems of mammalian,for machines.The functions include but not limited to information acquisition,processing,and understanding.Mathematically modeling of vision mechanism has high potential in providing scientific theory and guidance for computer vision researches.Recent years,biology and neuroscience have won great success in exploring the vision mechanisms.These researches provide a chance for computer vision community to model vision mechanism mathematically.The local structures of natural image are very complex,which contains rich information that is directly related to image semantic.Accurate representation of local image structure is therefore an essential challenge in computer vision.In this dissertation,some important visual mechanisms of mammalian are mathematically modeled to achieve more accurate representation of local image structure.The accurate representation of local image structure further solves some important problems in computer vision.In general,this dissertation makes the following contributions and innovations:1.Anisotropic mechanism of the primary visual cortex based local image structure representation: To model the anisotropic mechanism in the primary visual cortex,the pixel-wise anisotropy values of a given image are calculated by pseudo-Wigner-Ville distribution(PWVD)and Renyi entropy.Then the excitatory/inhibitory interactions among visual neurons in the local receptive field are modeled by measuring the similarities between their anisotropies.By mapping visual neurons to image pixels,the correlation between a central pixel and its local neighbors can be represented by a binary pattern which is named as local anisotropic pattern(LAP).Experimental results on texture classification and objective image quality assessment demonstrate that the proposed LAP accurately represents local image structures and further contributes to texture classification and objective image quality assessment.2.Local spiking pattern based local image structure representation: The neurons in primary visual cortex work by spiking pulses.A novel local spiking pattern(LSP)is proposed based on the mechanism of pulse spiking in the primary visual cortex.Through encoding the pulse-spiking results of the neurons in different iteration times,LSP obtains the relations between each neuron and its neighboring neurons.The extracted local relations are then utilized for representing the local structures of images.Experimental results on texture classification demonstrate that the proposed LSP represents the local image structures accurately.3.High-order local derivative pattern(LDP)based no-reference image quality assessment: Since conventional local binary pattern(LBP)is incapable of encoding high-order information and ignores the orientation of local image structure,this dissertation utilizes high-order local derivative pattern to extract quality-aware features and achieves non-reference image quality assessment based on the LDP features.The proposed algorithm uses Laplacian-of-Gaussian(LoG)to decomposition images into multi-scales in the first step.Each scale is then regarded as input to extract its LDP histogram.The LDP histogram is then utilized as features to train a regression model.Experimental results show that high-order local derivative pattern based no-reference image quality assessment achieves higher performance than state-of-the-art algorithms.