Research On Image Retrieval Based On Salient Region

With the developing of information technology, image has become a major information carrier. An increasing number of image data caused inconvenience to the manual processing and browsing. How to find interest image in massive data-base has become a problem need to immediately solved.Propelled by people's need content-based image retrieval is developed,which has a broad developing prospect.In this paper image retrieval technology based on salient region is discussed,which is a new technology in CBIR.The main content includes image salient region extraction,combination with multi-instance learning.In image Retrieval, usually as a result of the background image to occupy a large portion of the image, caused some impact to the feature extraction of the main objective of image .The salient region based retrieval approach is more conducive to content-based image retrieval. This paper first analyse salient region extraction approach based on salient map and wavelet salient point, then put forward the method of using the point as a clue, to find the salient region of the image, in the end color and texture features of the salient regions are extracted,and used for retrieval. Experiments prove the method has a great improve in precision rate than using overall features.Extraction of salient region of the image have improved the search results,but salient region is based on lower visual characteristics of images, often include the user's real region of interest, but the only way to find a region of interest is through machine learning .Traditional machine learning methods will feedback all the contents of a image as positive examples, the existence of a large number of pseudo samples have a bad impact to the learning outcomes.In multi-instance learning, the image is divided into many different instances, through learning, user's targe concept is acquired, which is in accord with that the user is often interested in some part of the image.In this paper,an approach combined multi-instance learning and salient region is put forward,which reuses the wavelet salient points,uses region around salient points as instance,based on the retrieval results,multi-instance learning is carried out,then resort the images according similarity.The application of multi-instance learning makes the results more consistent with searchers intent,so improves the retrieval precision.