Calvin Lab Auditorium
Today, the amount of data captured by sensors is outpacing our ability to process it efficiently. A common recourse is to partition the big datasets in smaller subsets and process them in parallel. However, such local processing approach will affect the underlying statistics, leaving unwanted artifacts, thus the need for a global approach.
Recent developments in algorithmic graph theory and linear solver technology have open up new tools such as combinatorial multigrid solvers. Image processing task are well suited to take advantage of the emergent graph techniques as images have underlying sparse graphs structure.
We would describe the transformation of fundamental image processing routines to solving grouped least squares problems over graphs. We would show how this global optimization approach gives better results over current processing techniques.
Next will be a discussion of implementations that leverages on new graph toolkits designed by the Big Data and Machine Learning community. Finally, results from tackling practical big image processing problems in remote sensing, using the graph-based approach, would be presented.