Some Issues On Interactive Image Retrieval Technical Content Of

With the explosive increase of the number of digital image, how to retrieve the useful and wantedinformation from the large database is an important problem. Content-based image retrieval (CBIR) arisesat1990s and is defined as retrieval relevant images from image dataset by extracting visual features.There are two types of descriptors for image content: global descriptor and local descriptor. Thecommonly used global descriptors include LBP (Local Binary Pattern), HOG (Histogram of OrientedGradients), color histogram and so on. Global descriptors is robust, thus is less vulnerable to noise.However, users are always looking for images that include specific objects, such as a car. Globaldescriptors fail to tackle this case, for a single signature computed for entire image can’t sufficientlycapture important properties of individual objects. Compared with global descriptors, local descriptors aremuch stronger for a single object.However, the image contains rich semantics which beyond the ability of visual features. How todeduce the semantic gap is an important research issue in CBIR. Relevance feedback (RF) is an effectivetechnology to deduce the gap. RF is first proposed and applied in information retrieval, then is introducedin CBIR. And it is proven that RF can boost the retrieval performance effectively. RF has four procedures:(1) The retrieval system returns the most similar images to user, this is the initial search;(2) The userindicates these images are positive or negative, these images are called feedback images;(3) The retrievalsystem learns user’s search intention by these feedback images;(4) Re-rank the database images and returnthem to user.First, our motivation is to propose an effective, real-time and robust RBIR technique, which is basedon learning region importance (RI) from user’s RF. In this paper, we have done following works.1. For reducing user’s burden of selecting region of interest, we construct a statistical index to describeRI. The RI index is based on human visual perception and intrinsic characteristics of images, and it also canbe learned from user’s feedback. Based on the proposed RI, an improved similarity measurement isproposed for region matching.2. In order to further satisfy users query intent, we propose a RF scheme with short-term learning, adaptive learning RI (ALRI). By increasing RI of positive images and decreasing that of negative ones,ALRI method raises importance of relevant regions and reduces that of irrelevant ones simultaneously.Extensive experiments on Corel-1000dataset and Caltech-256dataset demonstrate our proposedinteractive RBIR system is effective, robust and close to user intent.Second, this paper presents two content-based image retrieval strategies with RF based on particleswarm optimization (PSO). The first strategy exploits user indication of positive images. The second oneconsiders not only the positive but also the images indicated as negative. Both two RF strategies areimprovements of query point movement by assigning positive and negative images with different weights.These weights are learned by PSO algorithm. Experiments on Corel5000database show thecompetitiveness of our algorithm.