Luminance Channel Based Camera Model Identification

Camera model identification is an active research problem because of its importance in investigating the source and the authenticity of an image. Traditional camera model identification methods are based on strategies to extract the low-level traces left by the image acquisition pipeline of a camera on an image. One such intrinsic and camera-specific trace is the sensor pattern noise (SPN). The SPN is roughly approximated from the noise-residual obtained by performing high-pass filtering on an image. The noise-residual of an image also contains information about other types of noises. The extraction of the noise-residuals is generally performed on a single primary color channel, like the green channel of an image. However, the performance of a channel in the YCbCr color space is never explored. In this paper, we have proposed a novel camera model identification method based on convolutional neural network, where the noise-residuals are extracted from the luminance (Y) channel of the images. A constrained convolutional layer learns data-driven high-pass filters to extract the noise-residuals and the following layers learn a feature representation for the classification task. We have conducted experiments with multiple class combinations from the Dresden image database. The experimental results show the effectiveness of the Y channel for camera model identification both in terms of classification accuracy and convergence of the network.