Shadow detection in a single image

A method for generating a shadow mask of a digital image of a physical scene that identifies those areas of the digital image corresponding to physical elements in the physical scene that are in shadow, the method comprising the steps of:

• grouping adjacent pixels of the digital image into superpixels such that a superpixel comprises an equal number of pixels in height and width and such that each superpixel comprises approximately a same number of pixels;
• determining, for a first superpixel, both a largest eigenvalue and a next-largest eigenvalue given points in color-space defined by colors of constituent pixels of the first superpixel;
• generating, for the first superpixel, a ratio between the largest eigenvalue and the next-largest eigenvalue;
• repeating the determining and the generating for multiple other superpixels only from the digital image;
• normalizing, for each superpixel of the first superpixel and the multiple other superpixels, the generated ratio;
• and
• generating the shadow mask in accordance with the normalized ratio.

A system for processing digital images of physical scenes that comprise physical elements that are in shadow, the system comprising:

• a shadow mask generator for generating a shadow mask of a digital image from only that digital image, the generated shadow mask identifying those areas of the digital image corresponding to the physical elements that are in shadow;
• and
• an image modifier for modifying the digital image in accordance with the shadow mask;
• wherein the shadow mask generator comprises:
• a superpixel generator for grouping adjacent pixels of the digital image into superpixels such that a superpixel comprises an equal number of pixels in height and width and such that each superpixel comprises approximately a same number of pixels;
• an eigenvalue ratio generator for determining, for each superpixel, both a largest eigenvalue and a next-largest eigenvalue given points in color-space defined by colors of constituent pixels of each superpixel and for generating, for each superpixel, a ratio between the largest eigenvalue and the next-largest eigenvalue;
• and a normalizer for normalizing, for each superpixel, the generated ratio, wherein the shadow mask generator generates the shadow mask in accordance with the threshold value.