Research On Image Retrieval Based On Multiple Instance Learning

As traditional text based image retrieval needs textual annotation,it has some drawbacks:a large number of manpower,ambiguity,incomplete.With the emergence of the Internet,these problems are becoming more and more serious,which prompts the emergence of content-based image retrieval(CBIR).According to the low-level image features used in the retrieval process,existing content-based image retrieval(CBIR)methods can be categorized into two major classes,namely,global methods and localized methods.Global methods find query results by extracting and comparing their global,like color histograms,and color moment,etc.,with query images.These methods work very well when the major content of a query image is the interested object.However,when an object occupies a small part in the query image,these approaches can only include little useful information and the object's features are overridden by the background.As a result,localized CBIR appears.Multiple-instance learning(MIL)has been successfully utilized to address OBIR,where a bag corresponds to an image and an instance corresponds to a region of an image,as it mitigate the influence of background on the object by dividing an image into several instances.But existing methods cannot describe the localized features of image,namely interested object.In order to solve the above problem,in this paper,methods are studied respectively from the relationship among between the instances from different bags and the relationship between the instances in the same bag.In this paper we first present the research background and the present situation of image retrieval based on MIL,and analyzes the advantages and disadvantages of all algorithms We propose a new approach called multiple instance learning based on instance-consistency(MILIC)and the other called multiple instance learning based instances weighting and spatial relations(MILWS).The contributions are as follow:(1)For existing multiple instance learning methods cannot take full use of relationship among bags and instances,and instances and instances,we propose an algorithm,which is called multiple instance learning based on instance-consistency.We express relationship between the instances in different bags by instance-consistency,and propose a potential positive instance-selected method by instance-consistency and a feature map function based the potential instances to describe the interested object.Experimental results demonstrate the effectiveness of our proposal.(2)In order to describe the interested object more accurately and further improve sample' learning more accuracy in image retrieval,and use the relationship of instances from the same bag,we weight instances by the saliency of instances and express the spatial relations of instances by new spatial features.We propose a new approach called multiple based instances weighting and spatial relations.We design a potential positive instance-selected method based on weight of instances.Besides,we propose a new determine the number of potential positive instances by the weights of instances.Experimental results prove it can promotes the accuracy of the retrieval.