A Method to Detect Repeated Unknown Patterns in an Image

Consider a natural image that has been manipulated by copying, transforming and pasting back fragments of the image itself. Our goal is to detect such manipulations in the absence of any knowledge about the content of the repeated fragments or the transformations to which they might have been subject. The problem is non-trivial even in the absence of any transformations. For example, copy/paste of a textured fragment of a background can be difficult to detect even by visual inspection. Our approach to the problem is a two-step procedure. The first step consists in extracting features from the image. The second step explores the connection between image compression and complexity: a finite-context model is used to build a complexity map of the image features. Patterns that reappear, even in a somewhat modified form, are encoded with fewer bits, a fact that renders the detection of the repeated