Dual-branch convolutional neural network for robust camera model identification

Abstract

Identifying the model of the camera used for image capture is very useful in image forensics. To exploit complementary camera model specific features, we propose a dual-branch convolutional neural network (CNN) for camera model identification (CMI), where one branch directly uses the three-channel RGB image and the other uses a noise image obtained via high-pass filtering. For scalability, the approach operates on cropped image patches and majority voting is used for image-level CMI. We conducted extensive experiments to evaluate the proposed method on multiple datasets and compare its performance against prior approaches. For quantifying CMI accuracy, we use existing patch and image level (accuracy) metrics and also a new metric that we propose for assessing the robustness of image-level camera model estimates. Importantly, our evaluations and performance comparisons include: (a) datasets that are more representative of real-world application scenarios of current interest, where the images have undergone unknown processing as a result of sharing on social media platforms, and (b) cross-dataset scenarios where the evaluation is performed on a dataset different from and not necessarily represented by the training dataset. Our results demonstrate that the proposed approach offers improvements over the prior techniques, with particularly significant gains in accuracy for the social media dataset and for cross-dataset robustness assessment. The significant improvements over prior approaches that have used a single RGB or noise branch support our hypothesis that the proposed dual-branch architecture provides a convenient mechanism to introduce a favorable inductive bias in CNN architectures for CMI.