Research On Hierarchical Sparse Representation For Image Categorization

With the rapid development of Internet and multimedia technology,the images as the important carrier of information dissemination make human cognition of the objective world more simple.However,it has been an urgent problem how to efficiently analyze and manage information when we are faced with mass image data.Image categorization which has been regarded as a basic approach can be capable of addressing those problems and it has been widely applied in many scopes.How to train the computer to distinguish the inter-class similarity and intra-class variances has been a challenging task,in which extracting good image representations plays a vital role for image categorization.Therefore,this dissertation focuses on training strategies for deep architectures and modeling different hierarchical sparse representation.The major contribution and innovation include the following aspects:(1)With the development of deep learning in computer vision,feature representations have reduced reliance on hand-designed descriptors.Whereas the traditional dictionary learning performs sparse coding with spatial pyramid features based on scale invariant feature transform(SIFT)descriptor.For this purpose,a hierarchical learning architecture using automatically learned features from the pixel level is proposed,in which K-SVD with label consistency constraint is employed for training discriminative dictionary and optimal linear classifier.With regard to various benchmarks,the proposed algorithm only extracts densely sampled patches from gray or RGB images followed by hierarchical sparse coding.In view of the advantage of hierarchical features combined with supervised learning,the presented method significantly improves the classification accuracy.(2)Since previous research has shown that non-negative orthogonal matching pursuit(NOMP)as encoder is suboptimal in terms of computational cost,we study fast non-negative OMP(FNOMP)as an replacement and apply FNOMP into full-size image classification task based on hierarchical feature learning.The proposed method can accelerate utilizing QR factorization and iterations of coefficients.With the combination of FNOMP and unsupervised hierarchical training,we demonstrate that using gain-shape vector quantization for training dictionary,FNOMP performs not only more efficient than NOMP but also can significantly increase the classification accuracy than OMP based algorithm.Meanwhile,the experiments show that FNOMP based algorithm is superior to other state-of-the-art methods on several publicly available benchmarks.(3)In order to learn improved spatial pooling strategy in deep networks,a novel approach using learnable receptive field for hierarchical sparse representation is proposed.The pooling operator can be taken into account in a learnable framework where the pooling weights are jointly optimized together with a classifier.Hence,it allows for a richer set of possible pooling regions which lie on the data.Three different types of pre-pooling opertations including sum,average and max are analyzed followed by the discussion of sparsity level,the size of dictionary and receptive field,etc.The experimental results indicate that the presented method is effective for image classification task.(4)The main disadvantages of multiple kernel learning(MKL)based image categorization are the high computational complexity and conflicting results of effectiveness.To this end,hierarchical feature concatenation based kernel sparse representation is presented.Firstly,batch kernel OMP(BKOMP)is employed for single layer encoding on p.d.f gradients-based orientation histogram and spatial pyramid matching(SPM)features respectively.Next,the two types of low-dimensional kernelized sparse representations are concatenated as input for the second layer.Using kernel K-SVD(KKSVD)for dictionary training,the resulting kernel sparse codes are obtained with BKOMP.The experimental results show that the performance of low-dimensional kernel sparse representation significantly outperforms some MKL based approaches and other state-of-the-art methods.