Source camera identification based on an adaptive dual-branch fusion residual network

Although deep learning algorithms have addressed the issue of identifying the source camera to a certain extent, developing a straightforward and effective network remains a challenging task. At present, most of the excellent network schemes in source camera identification are deep networks, which heavily rely on the strong feature extraction ability of deep networks. Although deepening network layers has achieved certain results, training a deep convolutional neural network model requires a large dataset, sophisticated hardware and lengthy training time, and there is a waste of resources. To solve the problem of redundant structure and resource waste of deep convolutional neural networks, this paper proposes the SE-BRB module, which we call a new network module based on the residual module and SE module. Based on this, an adaptive dual-branch fusion network (ADF-Net) with a simplified structure is designed to identify the source of digital images. Specifically, the bottleneck residual module can achieve direct backward transfer of shallow features to avoid images being over-compressed and is suitable for capturing weak source features in images; Additionally, the introduction of a channel attention mechanism can increase the weight of effective feature channels in the network and improve network performance. Finally, multiscale camera feature fusion is realized through a dual-branch network structure to further improve the network performance. The accuracy of the model proposed in this paper is 99.33% and 98.78% on the Dresden dataset and the self-built complex dataset, respectively, and the classification accuracy is far ahead of the existing source camera identification methods.