Statistical Learning for Image Annotation

With the rapid expansion of computer networks and mobile devices, technologies are in great demand for managing and making best use of large collections of images. Semantic content analysis of images is at the heart of these technologies. The objective of this project is to develop new mathematical tools to tackle crucial and perplexing problems in image understanding. The statistical learning methods and the prototypical system to be developed in this project will have wide applications, valuable for both academia and industry. Software packages will be distributed and maintained for public access. This project will involve undergraduate students in research, and expose graduate students to a broad range of mathematical topics and interdisciplinary topics. Although advanced statistical learning methods have been exploited to annotate images automatically, existing approaches suffer from both flaws in the mathematical characterization of images and limitations in the modeling of annotation words and image components. In this project, new statistical models will be proposed for images and their semantics. Computationally efficient estimation methods will be developed for these models. Advanced optimization techniques will be used to better characterize images. New large-scale clustering algorithms will be developed for sets of weighted and unordered vectors under the Kantorovich-Wasserstein metric. It is envisioned that the clustering and statistical modeling methods developed in this project will have broader applications than image content analysis.