The Information Systems Team conducted trials and experiments in the area of content-based retrieval of texture images by addressing the problem of indexing these images based upon their feature vectors [2]. Image feature vectors usually require many dimensions (48) to store relevant information. Such long feature vectors are difficult to index and retrieve. The approach taken to mitigate this process involved truncating or compressing these feature vectors in order to obtain efficient performance in an R* tree, or similar multi- dimensional indexing structure.
Two types of processing, DFT (Discrete Fourier Transform) and SVD (Single Value Decomposition), were performed on each feature vector, to determine which provided better approximation to the original vector using the fewest number of vector dimensions. Using measures of recall and precision for epsilon and nearest-neighbor queries, it was determined that while SVD performs better overall, DFT is good for a small range of number of images suitable for browsing.