Abstract: Part I of this two-part paper proposed a robust way to detect local points on linear-isophote surface in an image. Only a subset of these points corresponds to linear surface in image irradiance and provides useful information about the camera response function (CRF). In Part II, we show that, for some images, this subset of linear points could constitute a very small portion of the candidate set and the remaining points are often considered as noise. Our previous approach was to eliminate the noise using a learning-based method. The learning-based method could only reduce the noise but not eliminate it completely. Hence, it fails when the proportion of linear points is too small. As a main contribution in Part II, we introduce the concept of edge profile and consider the candidate points as discrete samples of an edge profile. Instead of eliminating the unwanted candidate points as noise, we use them to instantiate the edge profiles. Assuming that every edge profile has a linear component in image irradiance, the interactions of the edge profiles in the space of linear geometric invariants may correspond to the linear part which is CRF-indicating. Such a model is shown to be sound and effective in both simulation images and real camera images.

@inproceedings{Ng:2009ac,
  author       = {Tian-Tsong Ng},
  url          = {http://www1.i2r.a-star.edu.sg/~ttng/papers/ng_wifs09_2.pdf},
  year         = {2009},
  pages        = {161--165},
  address      = {London, UK},
  title        = {Camera response function signature for digital forensics --- {P}art {II}: {S}ignature extraction},
  booktitle    = {Proceedings of the 2009 First IEEE International Workshop on Information Forensics and Security},
}