Content-based image retrieval using gradient projections

Content-based image retrieval (CBIR) enables a user to extract an image, based on a query, from a database containing a vast amount of pictures. This concept may be applied to many fields of interest including forensic science and image archiving. Current CBIR systems, however, are inaccurate. The purpose of this research project was to improve the accuracy of CBIR. The image's structural properties were examined to distinguish one image from another. By examining the specific gray level of an image, a gradient can be computed at each pixel. Pixels with a magnitude larger than the thresholds are assigned a value of 1. These binary digits are added across the horizontal, vertical, and diagonal directions to compute three projections. These vectors are then compared with the vectors of the image to be matched using the Euclidean distance formula. These numbers are then stored in a bookmark so that the image needs only be examined once. A program has been developed for Matlab on a Sun Sparc Computer with Unix Open Windows that performs this method of projecting gradients. Three databases were amassed for the testing of the proposed system's accuracy: 82 digital camera pictures, 1000 photographic images, and a set of object orientated photos. The program was tested with 100% accuracy with all submitted images to the database, and was able to distinguish between pictures that fooled previous CBIR engines. More importantly, though, was the program's ability to find certain similar scenarios in the database. This CBIR approach has significantly increased the accuracy in obtaining results for image retrieval.