Research On Automatic Image Annotation

With the constant progress of multimedia technology and the rapid development of computer network technology,how to effectively manage and retrieve images become an urgent demand.Automatic image annotation plays an important role in the semantic-based image retrieval.However,during the process of automatic image annotation,there still exist some problems needed to be further solved: the feature extraction of images;the correlation between the labels;the ambiguity of complicated labels.This thesis mainly concentrates on solving the above-mentioned problem.The main work of this thesis is conducted as follows:(1)To alleviate the feature extraction of images and correlation between the labels these two problems,an automatic image annotation algorithm based on tensor penalized partial least square(Tensor-PPLS)is proposed.First,an image is expressed as a tensor according to the low visual feature and tensor structure of the image.Tensor structure can keep the correlation between pixels.Then,multilinear principal component analysis of tensor objects(MPCA)is used for dimension reduction,removing redundant information.Finally,we propose TensorPPLS algorithm to annotate images.Tensor-PPLS can maintain the correlation between the labels,considering the sparse label at the same time.The comparative experimental results on the Corel5 K,IAPR TC-12 and MIRFlicker benchmark image datasets illustrate that the proposed algorithm can effectively perform image annotation tasks.Tensor structure of the image is helpful for image annotation.(2)An automatic image annotation algorithm based on covariance fast multi-instance multi-label(C-fastMIML)is proposed,which takes into account the above three questions.First,an image is segmented into several regions,every region is expressed by a covariance matrix.Covariance matrix proposes a way of fusing multiple feature,describing the information of objects in the image well.Furthermore,covariance matrix is low-dimensional compared to other region descriptors.Then,based on MIML,we propose C-fastMIML algorithm to annotate images.The correlation between the labels and ambiguity of complicated labels can be solved in C-fastMIML algorithm.The performance of the proposed algorithm is tested on Corel5 K and IAPR TC-12 benchmark image datasets.The results show that the proposed algorithm is effective and efficient compared with other state-of-the-art automatic image annotation algorithms.