A parallel algorithm for multi-feature content-based multimedia retrieval Punpiti Piamsa-nga, Nikitas A. Alexandridis, Sanan Srakaew and George Blankenship EECS, George Washington University, Washington D.C. 20052, USA {punpiti, alexan, srakaew, blankeng}@seas.gwu.edu ABSTRACT In this paper, we propose a parallel, unified model for the indexing and retrieval by content of multimedia data using a weighted cascade. [4] Each multimedia data type can be viewed as k-dimensional (k-d) data in spatio- temporal domain. [3] Each dimension of the data is separated into small blocks and then formed into a multidimensional tree structure, called a k-tree. Using the k- tree structure, the retrieval time improves while the retrieval accuracy remains relatively constant. Moreover, since we can realize all types of multimedia data using the same k- tree data structure, the data indexing and retrieval algorithms are uniform. In this paper, we demonstrate a parallel algorithm with a weighted cascade that can be used with the uniform model. The parallel processing and weighted cascade algorithm improve the retrieval time. We evaluated the performance results of multimedia database queries using a Beowulf-class cluster of workstations. [1] The experimental results indicate that the retrieval accuracy improves with the utilized tree depth, while the parallel processing minimizes the retrieval time. Keywords: k-tree model, multidimensional signals, similarity search, multimedia data retrieval, weighted cascade, and parallel processing. 1 INTRODUCTION Content-based indexing has become more important since conventional databases cannot provide the necessary efficiency and performance. [2] However, it encounters three major difficulties. First, the data content (or feature) is subjective information that is used to characterize the data. The recognition of data content requires prior knowledge and special techniques in Computer Vision and Pattern Recognition. Second, if a method or processing technique is designed and developed for one type of data or feature, it may not be appropriate for others. For example, a technique designed for indexing audio data may not be usable for image data; or, a technique developed for a color feature may not be appropriate for a texture feature in image and video data. Third, the huge data size and the requirement for a similarity search affect the computation. Similarity matching requires the computation of the distances between a query and all records in the database; the best match is the data set with has smallest distances. To solve these three problems, we use a mathematical model to represent features; a k-tree model to represent the data structures of the multimedia data; and exploit parallelism to reduce the retrieval time. In this paper, color and texture are the features of interest; they represent the subjective information of the multimedia data. We use a normalization technique to generate the indices. [4] The domain of a feature is reduced to a set of selected values from a universe of potential values for the feature. We use an identification number of each element in the reduced set. When data is inserted into the system, it is converted to the selected domain. The feature is represented by a histogram. For color feature, a few colors are picked from the whole infinite universe of colors. Each color is indexed by a finite number. The color feature of an image or a video is represented by a histogram using the indexed color. The comparison of two features is based upon the distance between the histograms that define the features. A k-tree model [4] for multimedia data is used to unify the data characteristics. A k-tree is a directed graph; each node has 2 k incoming edges and one outgoing edge with a balanced structure. A k-tree is a binary tree for 1- dimentional data; a quadtree for 2-dimension data; and an octtree for 3-dimensional data. There are three main benefits for exploiting a k-tree. First, the spatio-temporal information of the data is embedded into the tree structure. Therefore, it reduces the time to compute distances between two nodes when spatio-temporal information is required in a query. Second, a k-tree can exploit multiresolution processing by computing small, global information first and then large, local information when more accurate resolution is required. Third, the complexity of data structure affects only the degree of the tree. Consequently, an algorithm for a particular type of feature can be reused for a feature of other media types To reduce the response time, we proposed two techniques in our previous reports, one is a weighted cascade (pipeline) for multiple-feature querying [4] and the other is to exploit parallelism during the search. [5] Both techniques separately demonstrate a reduction of the retrieval time of an image from a 30,000-record image database. In this paper, we exploit parallelism with a unified k- tree model for multimedia retrieval; the model produces multiresolution processing of a unified structure that is able to accomplish multi-feature queries. We apply temporal and spatial parallelism using a network of workstations. We also present the performance measurements for a multiple-