Computer Science Colloquium, 2006-2007

Statistics in Hybrid Imaging

Computer vision typically regards images as given entities to be processed. However, richer information is gained by modifying and analyzing the imaging process itself. Such "hybrid imaging" exploits the power to affect both the sensor and the algorithmic components of a vision system. As it turns, the methods involved have interesting statistical considerations.

We review several of our results in this context. We handle needs of imaging objects under various illumination settings: there, multiplex codes greatly improve image quality, by accounting for real image noise characteristics. Then, we look at imaging in scattering media. We recover visibility and even a 3D structure of hazy outdoor scenes (e.g., Haifa Bay or Scuba diving) by compounding a physical model with independent component analysis (ICA). Here, steps are taken for efficiently optimizing entropy and mutual information, of general scope. Finally, we explore cross-modal sensing and statistical analysis, by looking at Audio-Visual interaction as a case study. Data insufficiency is overcome by a general mathematical principle, based on a sparsity prior. It uniquely distinguishes moving image features associated with sound sources.