Partial Highest Possible Edge Analysis for Interactive Image Accessibility

Abstract
Relevance feedback is a technique that takes advantage of human-computer interaction to refine high level queries represented by low level features. Among RF schemes, the most popular technique is SVM based RF scheme. When SVM is used as a classifier in RF, there are two strategies. One strategy is to display the most positive images and use them as the training samples. The most-positive images are chosen as the ones farthest from the boundary on the positive side, plus those nearest from the boundary on the negative side if necessary. Another strategy is that most of SVM based RF scheme does not consider the unlabeled samples even though they are useful in constructing a good classifier. To overcome these drawbacks, in this paper we propose a biased maximum margin analysis (BMMA) and semi supervised BMMA (semiBMMA) for integrating the distinct properties of feedbacks and also to utilize the information of unlabeled samples. The BMMA differentiates positive from negative feedbacks, whereas semiBMMA takes into account the information of unlabeled samples by the introduction of Laplacian regularizer to BMMA. To validate the efficacy of the proposed approach, we test it on both synthesized data and real-world images. Promising results are achieved and this can significantly improve the performance of CBIR systems.
Keywords: Content-based image retrieval (CBIR), relevance feedback (RF), support vector machines (SVM),graph embedding framework.
I.Introduction
Content-based image retrieval (CBIR), as we see it today, is any technology that in principle helps organize digital picture archives by their visual content. By this definition, anything ranging from an image similarity function to a robust image annotation engine falls under the purview of CBIR. This characterization of CBIR as a field of study places it at a unique juncture within the scientific community. In the CBIR context, an image is represented by a set of low-level visual features, which are generally not effective and efficient in representing the image contents, and they also have no direct correlation with high-level semantic information. The gap between high-level information and low-level features is the fundamental difficulty that hinders the improvement of the image retrieval accuracy. Recently, a variety of solutions have been suggested that aim to bridge this semantic gap. The relevance feedback [1] narrows the semantic gap by making use of user provided judgments which are the labels (relevant or non-relevant) on the retrieved images for a query. The retrieval performance improves as the user provides more and more feedback information to the CBIR system. Query vector modification (QVM) [2] and feature relevance learning [3] are the two widely used methods to integrate user feedback information into the CBIR system. Majority of the work uses relevance feedback to learn the relative importance of different features, with some tries to learn a feature weighting scheme either with [4] or without[5] considering correlations among feature components; while others either use a probabilistic scheme , or Self-Organizing Maps , or boosting technique , etc., to do so. A typical problem with CBIR system with relevance feedback is the relatively small number of training samples and the high dimension of the feature space. The system can only present the user with a few dozen of images to label (relevant or irrelevant). The interesting images to the user are only a very small portion of the large image database, in which most images remain unlabeled. Therefore, small sample learning methods are most promising for RF. Two-class SVM is one of the popular small sample learning methods widely used in recent years and obtains the state-of-the-art performance in classification for its good generalization ability. Guo et al. developed a constraint similarity measure for image retrieval which learns a boundary that divides the images into two groups, and samples inside the boundary are ranked by their Euclidean distance to the query image. The SVM active learning method selects samples close to the boundary as the most informative samples for the user to label. It is almost impossible to estimate the real distribution of negative images in the database based on the relevant feedback. Nevertheless most of the SVM RF approaches ignore the basic difference between the two distinct groups of feedbacks, i.e all positive feedbacks share a similar concept (Fig1) while the negative feedbacks share a different concepts (Fig 2). Directly using SVM as an RF scheme damages the entire performance of CBIR systems. One problem is that different semantic concepts live in different subspace and it is the goal of RF schemes to figure out “which one”. Additionally it has another problem of incorporating the unlabeled samples into traditional SVM based RF schemes, even though they are useful in constructing a good classifier. To explore solutions to the above problems, in this paper we propose a technology, biased maximum margin analysis (BMMA) and semi supervised BMMA (SemiBMMA) for the traditional RF scheme, based on graph embedding framework [30]. The proposed scheme is mainly based on the following: 1) effectiveness of treating positive and negative samples differently; 2) the success of graph embedding in characterizing intrinsic geometric properties of the data set in high dimensional space; The convenience of graph embedding framework in constructing semi supervised learning techniques.