Thomas Deselaers: Reducing time and RAM Requirements in content-based image retrieval using retrieval filtering

In this paper we present and evaluate how computationally cheap comparison measures can be applied in content-based image retrieval applications to reduce the timeand memory requirements. The time requirements are reduced by applying filtering techniques and evaluating computationally costly distance measurements only to a suitably chosen subset of images. The RAM requirements are reduced by keeping only those features in memory that are absolutely required. It is shown that runtimeand memory efficiency is greatly improved with hardly any changes in retrieval quality. Introduction With the ubiquity of cameras and the ever-increasing necessity of digital images in medicine, the amount of images stored in databases is growing quickly. Access to these data is commonly achieved using textual meta data. Content-Based Image Retrieval (CBIR) systems are an alternative and allow accessing image databases by image content rather than by textual information. A problem with CBIR systems is that they require computationally expensive operations and large amounts of memory to allow for acceptable results. The aim of a CBIR system is to find visually similar images for a given query image. To find visually similar images, typically features are extracted from each image in the database and from a given query image. Then, the features of the query image are compared to the features of each database image and thus the most similar images from the database can be determined. Some of these distance measures, e.g. the Image Distortion Model (IDM) [KGN04], provide good results in terms of error rates but have high computational costs which do not allow for interactive use in the context of large databases. The concept of filtered retrieval, i.e. use a computationally cheap distance function to preselect images for the computationally more costly distance function, is wellknown in the database and data exploration community, e.g. [FBF94] and [SH94] propose to use a lower dimensional distance function as a filter for a higher dimensional quadratic distance function. Filtered Retrieval in FIRE We integrated the concept of filtered retrieval into the Flexible Image Retrieval Engine (FIRE) 1 [DKN04]. In FIRE, an image X is a set of features X1, . . . , XM representing certain aspects of the images. Normally, FIRE calculates the distance between a query imageQ and a given imageX contained in the database as a weighted sum over all individual feature distances. This calculated distance is used to assign a similarity score S (Q,X) to 1http://www-i6.informatik.rwth-aachen.de/∼deselaers/fire.html