Interactive Visual Analysis of Images

Mining and visualizing huge image databases has become a daunting task for many application areas such as astronomy, medicine, geology, oceanography, and crime prevention. In this thesis, we introduce a new technique for feature extraction that obtains image signatures from pixels obtained by quantized images at different levels. We also present an application (PixSearcher) that retrieves and organizes huge images databases. In contrast to raster techniques that use the entire pixel raster for distance computations, our application uses a small set of descriptors to handle large image collections. Our research program began in 2010, when we were involved with characterizing the 2D distributions of orthogonal pairwise projections on a set of points in multidimensional Euclidean space. These characterizations included measures such as density, skewness, shape, outliers, and texture. Working directly with these measures, we located anomalous or interesting distributions in large collections of pairwise projections. For example, we used these measures to locate unusual correlations between pairs of time series and we used them to reorganize scatterplot matrices to avoid clutter. Although the original motivation was to characterize the 2D distributions of scatterplots, we soon realized the idea had more general implications. We extended this work to handle pixels in images and developed new descriptors that are appropriate for images. Using these descriptors led to algorithms that outperformed conventional techniques on speed and accuracy.