In the context of images, the classic reference is the book "Vision" by David Marr, in which the idea of zero-crossings is heavily exploited in his theory (now, I think, largely superseded) of early visual processing in animals. Any reasonable academic library should have a copy of this seminal text.
Images do not have zero-crossings; however if you convolve an image with a differencing operator such as a Laplacian of Gaussian (or Difference of Gaussians) the result will usually have zero-crossings. These are simply the zero-level contours between positive and negative regions of the array; their positions may be estimated in a discrete array by interpolation, or crudely approximated by finding pixels which have an opposite-sign neighbour.
If you have the Image Processing Toolbox, then the 'log' option of the edge function incorporates an implementation of zero-crossing detection.
