Effective and efficient region-based image retrieval

Abstract

Content-based image retrieval (CBIR) is a challenging task. Current research works attempt to obtain and use the semantics of image to perform better retrieval. Towards this goal, segmentation of an image into regions has been used in recent years, since local properties of regions can help matching objects between images and thereby contribute towards a more effective CBIR. This paper improves on a CBIR technique, called SNL (Sridhar, Nascimento, Li) that utilizes the regional properties of the images. In SNL each image is segmented and features including the color, shape, size and spatial position of the obtained regions are extracted. Regions are then compared using the integrated region matching (IRM) distance measure, which is not a metric, which prevents the use of metric access structures or filtering techniques based on the triangle inequality. We overcome this issue, by using MiCRoM, a true metric distance to compare segmented images. This resulting approach, called SNL ∗ , can be used in conjunction with a filtering technique to reduce substantially the number of images compared. Albeit metric-based, SNL ∗ is computationally expensive. We address this drawback, in the SNL + approach, where we replace the expensive metric distance in SNL ∗ by the inexpensive original (non-metric) IRM distance. We found that one can still make use of the same filtering technique, at the expense of little loss in retrieval effectiveness. Thus, the main contribution of this paper is SNL +, a very effective and highly efficient region-based image retrieval technique.