Source Camera Identification Method Based on Multi-Scale Feature Fusion

Source camera identification has become one of the research hotspots in the field of digital image forensics. Most existing methods are based on deep neural network models. While these methods have improved traceability accuracy compared to traditional methods, their performance in terms of accuracy becomes mediocre when the device cat-egories are expanded. Moreover, they require retraining on a complete dataset or fine-tuning on newly added datasets. To address these challenges, this paper introduces a source camera identification method based on multi-scale feature fusion. Different scales of convolutional kernels are used to sample the input image, and parallel residual networks obtain sensor pattern noises at different granularities. A fusion network layer then inputs the merged features into a Softmax layer for classification results. Furthermore, to avoid repeated training due to class expansion, high-dimensional network features are extracted to construct an index vector database for retrieval classification. Experimental results demonstrate that the multi-scale feature fusion method achieves higher accuracy in camera traceability tasks. Additionally, the proposed retrieval mode effectively addresses the category expansion problem with minimal accuracy loss. © 2024, Taiwan Ubiquitous Information CO LTD. All rights reserved.