Understanding visual complexity with statistical physics

An image of natural scene may be described in a few words, yet an image of static requires describing all of the pixel values. We call this property of images visual complexity. Different perspectives for understanding why natural images have low visual complexity have wide applications in fields like computer vision and image compression. This REU project reviews conventional approaches to visual complexity related to the field of natural image statistics, and proposes two new perspectives on visual complexity which do not require such statistics: a notion of information for an statistical ensemble created by modifying an image, and a notion of how informative different coordinate systems are for describing points from a probability distribution. Taken together, these different perspectives form a framework for understanding visual complexity in a variety of images