There are many reasons why someone might wish to know the precise camera that was used to take a digital photo – whether for criminal or fraud investigation, copyright and provenance, and perhaps even for archival purposes. Work published in the International Journal of Computational Vision and Robotics, provides a novel feature-based approach for such an identification using photo-response non-uniformity (PRNU) noise.
Megha Borole and Satish Kolhe of the School of Computer Sciences at Kavayitri Bahinabai Chaudhari North Maharashtra University in Jalgaon, Maharashtra, India, explain how the pattern of noise in a digital image can act as a “fingerprint” unique to a particular camera. It can even be used to distinguish between the same make and model of camera with the same lens. “PRNU noise exhibits a different noise pattern for each image sensor and if numerous pictures are taken of a similar scene it remains around same,” the team explains.
The team explains that, somewhat paradoxically, they begin by applying a “denoising” procedure to the digital photo of interest. The filter allows them to reveal the PRNU noise pattern. This output is distinct from generic photographic noise and is, the team explains represented by the pixel intensities known as the Hu set of invariant moments. These invariants persist under image scaling, translation, and rotation, unlike many other characteristics of a digital photograph which may be lost when the photo is manipulated. The next step is to feed these features into a fuzzy min-max neural network (FMNN) that has been trained and classified with known digital cameras beforehand.
The team has demonstrated proof of principle for the approach with seven camera groups and showed that they could identify the specific camera used to take a photo of the same scene as all the others more than nine times out of ten on average. Given that in any real-world situation there may well be other evidence to point to a specific camera in many kinds of investigation where its identity needs to be known. The next step will be to improve the behaviour of the neural network by reducing the impact of inherent random noise.
Borole, M. and Kolhe, S.R. (2021) ‘A feature-based approach for digital camera identification using photo-response non-uniformity noise’, Int. J. Computational Vision and Robotics, Vol. 11, No. 4, pp.374–384.