Research On Image Retrieval And Classification Method With Visual Descriptors

With the rapid popularization of personal media devices and the emergence of the internet technology,the number of images in real life grows rapidly,so how to organize and manage these images quickly and effectively becomes a hotspot of research.In addition,the rapid development on all kinds of emerging technologies has leaded to the unceasing change of users' requirement for image retrieval and classification,especially in the demand to understand semantics of the image,the fact that the content of images becomes increasingly rich intensifies the “semantic gap” between underlying visual description information and high-level semantic visual description information,which has brought a great challenge to the image retrieval and classification technology.Images own wealthy visual description information including color,shape,texture,the local,spatial information,and the high-level text information descriptors,etc.These features used to explain the image visual description information are collectively known as visual descriptors.Visual descriptors have very important significance for image retrieval and classification.Around the image visual descriptor representation,extraction and the use,this paper mainly focuses on the image retrieval and classification to study,whose main works of this paper are embodied in the following two aspects:(1)Put forward an image retrieval method based on rough set inference rules and color mutual information descriptor.In order to decrease the restriction which is brought by the uncertain information to the results of image retrieval,based on the mapping from underlying descriptor to high-level semantic having been completed and the text representation having been formed for images,according to the theory of rules extraction in rough set,an image retrieval model based on rough set inference rules is proposed.However,in this model,the mapping from low-level visual descriptors to high-level semantic descriptors exists information loss,which causes the “semantic gap” problem.In order to release the retrieving imprecision brought by “semantic gap” problem,after the image retrieval model based on rough set inference rules having been completed,for one more step the related visual descriptor of color is introduced,and then the image retrieval method based on rough set inference rules and color mutual information descriptor is proposed,which takes full consideration of the advantages in underlying visual description variation and realizes the precise retrieval of information.Through the integration of the proposed two image retrieval models,a multi-model image retrieval scheme is realized.The experimental results show that the proposed method can improve the precision of image retrieval effectively.(2)Put forward an image classification method based on spatial difference descriptor.In the image classification field,a famous model is spatial pyramid matching model.In the aspect of feature encoding,the spatial pyramid matching model is an extension of the Bag-of-Words.Firstly,it divides an image into gradually growing sub areas,and then cascades all sub area histograms.However,the spatial pyramid matching model does not consider spatial difference information between sub areas.In order to make full use of the ignored information,a new descriptor is put forward,called spatial difference descriptor,and then the image classification method based on the spatial difference descriptor is proposed in this paper.This method mainly contains five steps: low level descriptors extraction,sparse coding,descriptors pooling,spatial difference descriptors computation and linear classification,among which the spatial difference descriptor computed out is used to connect histograms of Bag-of-Words model in spatial pyramid matching model.In order to verify the effectiveness of the proposed method,experiments on Scene15,Caltech101 and Caltech256 database are conducted in this paper,and the experimental results show that this proposed method improves the accuracy of image classification effectively.