Robust camera model identification using demosaicing residual features

Abstract

In this paper, we propose a new framework for performing accurate and robust camera model identification by fully exploiting demosaicing information in a camera’s output images. Instead of fitting a camera’s demosaicing process into parametric models, our framework works by exposing and extracting a diverse set of intra-channel and inter-channel color value correlations originated from the demosaicing process. To expose these correlations, we first apply a number of diversified baseline demosaicing algorithms to re-demosaic the image under investigation, and gather a set of both linear and nonlinear demosaicing residuals. To further extract demosaicing correlations with respect to the color filter array (CFA) structure, co-occurrence matrices are calculated using a new set of geometric patterns. These patterns are specifically designed to extract different types of color value dependencies within the repeated lattice of the CFA pattern. We design a multi-class ensemble classifier to utilize all extracted color value correlations to perform camera model identification. A series of experiments show that our proposed framework can achieve an accuracy of 98.14% on a database with 68 camera models, and is highly robust to post-JPEG compression and contrast enhancement.

Appendix

In this appendix, we present the convolution filters designed to realize the horizontal and vertical bilinear demosaicing algorithms in our baseline algorithm set.

In Fig. 12, we show the red, green and blue layers after re-sampling according to the Bayer pattern. The blank blocks correspond to missing color components that must be interpolated using the sampled colors. Due to the symmetry of the green layer, all the missing green components can be interpolated using the same filter. The filter coefficients used to interpolate the green components are [0.5, 0, 0.5] for the horizontal bilinear demosaicing algorithm. For the vertical bilinear demosaicing algorithm, the filter is the transpose of the horizontal one, i.e. [0.5, 0, 0.5]^T. For convenience, we use horizontal filter to refer to the interpolation filter for the horizontal bilinear demosaicing algorithm, and similarly, vertical filter for the vertical bilinear demosaicing algorithm.

Fig. 12

Red, green and blue layer after re-sampling according to the Bayer pattern

For the red and blue layers, the three color components to be re-demosaiced in each Bayer pattern require different filters. For colors in position P1 in both the red and blue layers, the horizontal filter is [0.5, 0, 0.5] and the vertical filter covers a 5 × 3 window:

$$ \frac{1}{6}\left[ \begin{array}{lll} 1 & 0 & 1 \\ 0 & 0 & 0 \\ 1 & 0 & 1 \\ 0 & 0 & 0 \\ 1 & 0 & 1 \end{array} \right] $$

(10)

For missing red and blue components in position P2, the horizontal filter is:

$$ \frac{1}{2}\left[ \begin{array}{lll} 1 & 0 & 0 \\ 0 & 0 & 0 \\ 0 & 0 & 1 \end{array} \right] $$

(11)

and the vertical filter is the transpose of the horizontal filter. For color components in position P3, the vertical filter is [0.5, 0, 0.5]^T and the horizontal filter is:

$$ \frac{1}{6}\left[ \begin{array}{lllll} 1 & 0 & 1 & 0 & 1 \\ 0 & 0 & 0 & 0 & 0\\ 1 & 0 & 1 & 0 & 1 \end{array} \right] $$

(12)

which is the transpose of vertical filter for color components in position P1.

We can see that every time a color component is interpolated, the horizontal demosaicing algorithm uses more re-sampled colors in the horizontal direction and the vertical demosaicing algorithm has a preference for colors in vertical direction. By designing the interpolation filters in this way, we expect vertical and horizontal bilinear demosaicing algorithms to expose directional information of a camera’s demosaicing process.