Image Retrieval Method Research Based On Granular Computing

With the rapid development of image acquisition technology,sensing technology and Internet technology,the number of digital images in real life has increased dramatically.Facing with the image database which has huge number and rich content,how to find the images that people is interested in accurately and quickly has become increasingly important.As a theory,technology and tool to solve complex problem from multilevel and multi-angle in rencent years,granular computing has been widely used in data mining,uncertainty information processing,image processing,and other fields.For the problems in content-based image retrieval,by analyzing the shortcomings and the insufficiency of existing retrieval methods,the ideas and methods of granular computing are introduced into the process of image retrieval.From the aspects of feature representation,feature indexing in image retrieval,two image retrieval methods based on granular computing are proposed.The main contents of this paper are as follows:(1)Aiming at the problem that the typical feature indexing methods are based on image low-level feature,which has a “semantic gap” with human's understand to image content,an image retrieval method based on multi-granularity partition is proposed.Firstly,the image information table is established according to the semantic features of the image.The rule set is extracted by the multi-granularity rule acquisition algorithm,and the semantic feature index is established according to the knowledge particle corresponding to the rule set.Secondly,a similarity metric based on connotation importance is used to measure the similarity between the images.Then,the most similar image set is obtained by the rule matching,the similarity between the images in the candidate set and the query image is calculated,which avoid querying from the whole image database;Finally,through the experiments on the Corel database proved that this method can ensure the retrieval accuracy while improving the retrieval efficiency.(2)The common feature pooling methods based on bag-of-words(BoW)based image retrieval method are easy to lose the feature diversity.In view of this problem,the cloud model is introduced into the BoW-based image retrieval process,a feature pooling method based on cloud model is proposed,and a more discriminative image feature representation is obtained,on this basis,an image retrieval method is proposed to preserve the diversity of BoW features.Firstly,the SIFT feature of the image is extracted and the visual word codebook(visual dictionary)is constructed.Secondly,for each image,all the SIFT features in the image are mapped into the corresponding visual word.Then,a feature pooling method based on cloud model is used to generate a feature vector which has large amount of information,and the feature vectors are used to calculate the similarity.Finally,the experiments on the Corel and Holidays databases show that the method can obtain more discriminative image feature expression.